I think this is the first time we've had a third minor version bump on a frontier Anthropic model. (I count the 0.5s as major here, because they've been issued non-sequentially and also corresponded to massive capability leaps, eg, Sonnet 3.5, Opus 4.5).
So now the Opus 4.5 family has successors 4.6, 4.7, and 4.8, each posting fairly modest claimed gains. My own experience w/ 4.6 and 4.7 are that I don't firmly grasp any capabilities improvements over my memory of 4.5, but it's all so fuzzy that it's truly difficult to tell.
Maybe my own tastes are saturated now (it's smarter than me?) and I'll never again perceive model progress. Maybe the incrementalism is such that I'd notice immediately if my 4.7 workflows were redirected now to 4.5.
Difficult spot for the labs to be in because, if they have a stronger product, I'd prefer they release it and that I can use it.
But as this dynamic continues, the improvements are going to be less and less legible for end-users, who will complain about the churn-without-payoff, even when the payoff may actually be real.
I won't be surprised if the next gen frontier models are the last.
There's orders of magnitude of low hanging juice to squeeze out of smaller models.
It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years (design not certain, probably unlikely).
It is far less clear that a 1.2T model will be meaningfully better enough to justify training it.
As far as reasoning is concerned, with the recent GRAM release, there may be 4 orders of magnitude of reasoning to tack on to smaller models.
Think about that... Google, OpenAI, Anthropic could train a 30B GRAM-based model in days - and it could potentially have better local reasoning than the best model available today at >1T params... They could upgrade that to a ~600B MoE model in days to have general trivia knowledge rivaling the best models...
You just can't train a 1T+ parameter model that fast. It is a giant if how much GRAM turns out to improve things, but it's unlikely to be trivial or nothing.
Larger models can already sort of tell you anything. They're never going to get everything right unless they stop being LLMs.
There's just not a lot of juice left to squeeze for Gemini to tell you exactly how tall Ke$ha is or when the last time Brittney Spears went to jail was...
>It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years.
I don't disagree, but how much of this ends up being distillation? I can't help but imagine that 4.8 was probably trained in part by leveraging Mythos.
If the very large models turn out to be very expensive to run relative to the benefits, it's possible that they could end up still being trained, but ultimately used as a tool to create smaller models that are nearly as effective.
I'm curious if someone here with a stronger background in the space has a similar intuition or not.
It’s really worth distinguishing between old fashioned student teacher distillation and large scale synthetic dataset creation.
The latter is much better (since you can clean up, review, update responses and filter your datasets).
I suspect nobody is doing real student teacher distillation, it’s just easier to do a bunch of training on the same giant corpus then fine tune via the synthetic corpus (which might have been generated by a bigger better LLM)
> I don't disagree, but how much of this ends up being distillation?
A lot, so you can bet tens of millions are flowing to congress to have distillation declared illegal before this happens. And then it'll happen anyway.
A lab can train a large model, and then distill a smaller model from it that retains the majority of the useful capbility.
I don't know well enough if there's any benefit of that over just training the smaller model directly, but I'll bet there are some times where that is useful. I could easily see it being easier to do the initial pre-training on a larger model but be able to distill everything useful down into a smaller model, essentially filtering out a lot of noise in the process.
There used to be training methods like that but I think they've been phased out in favor of letting small models evolve by rewriting their own training material. Surprisingly that's actually cheaper.
> I don't disagree, but how much of this ends up being distillation?
You don't need distillation. They already have the training sets.
It's MLA + MoE + Medusa (a better version of Speculative Decoding) + 1.58b (possibly - maybe nothing) + GRAM (which will almost certainly not turn out to be a nothing burger, but no one has quickly turned this around yet to prove it).
The frontier labs distill their own base models all day long. It’s not just something done by nefarious Chinese copycats. The knowledge embodied by the internal base models that we never see is much more powerful and useful than the much sparser raw training data
It wouldn't be data distillation: instead, it would be teacher-student distillation. The teacher model has stronger representations that the student can mimic, which would give it more capability over training on the data itself.
Frontier labs have their own variants of MLA and certainly their own balance/scaling-laws for things like MoE vs FC vs Attn. MoE scales really well for inference with horizontal scaling + batching, which these guys luv.
On the architectures side, I'm a lot more interesting in attention residuals than anything else, one of those things that seems obvious in hindsight and Kimi have proven it at scale.
Same with speculative decoding... They all do something, but there are known techniques that are substantially better - that just were't known when they started development of the previous models.
How useful is speculative decoding in a batched setting where you get paid for throughput (aggregated across users) and you mostly don’t get paid for latency or single-session throughput?
I looked into this "GRAM" stuff a sibling comment links further to, and just to say:
- this gets reinvented/rediscovered constantly under different names
- it cant be trained very well (right now, will change)
- massive theoretical improvements over current models (log_2(vocabsize)=17, residual stream dim is thousands of dimensions, recursivity means more information bandwidth by ~3 OoM)
- BUT it cant be interpreted or aligned <- this is why no one uses it and no one talks about it. the idea is 100% obvious to all the frontier labs and there is a good reason why it isn't used
I follow this stuff closely, I think I know what I'm talking about
What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.
Honestly, there is nothing in my head that Claude cannot handle. Maybe it can be more this or that but I can already barely exploit Opus 4.7.
And I'm using DeepSeek 4 Pro for my personal use and while it's a little behind, it's not that far.
I think the situation can be very dangerous for US AI companies because if current models are already capable of doing mostly anything, nobodoy will want to get to the next model, even if it's 10x better. OTOH, open source models like DeepSeek are doing mostly the same work for 1/10 of the price.
Also the more I play with Pi, the more I think LLMs are already not kept back by their own capabilities but by the lack of agency we allow them to have. There is more value today in a capable harness for current LLMs than in a better LLM.
>What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.
I think what gp said was the improvements are incremental, and we haven't seen a big revolutionary change since 2-3 years, and the pace is slowing down.
Have you personally used any of the latest batch of even smaller local models? They certainly don't beat SotA models at coding... but with a good harness they are able to achieve things with SotA that I couldn't last year.
I've repeatedly given local models non-trivial projects that involve research and coding which they've successfully completed with minimal intervention from me (almost exclusively in the domain of reviewing the results). Again, nothing comparable with current SotA, but definitely tasks I could not have given SotA models last year (without agent harness).
Now that pure progress from these models seems to have slowed down, we're seeing a ton of options for both making models more efficient and other tools that help improve them (everything from agent harnesses to RLVR).
That's just looking at "what can small do today", when you look at what's possible with larger open models that are still much smaller than SotA from the major providers, their performance is extremely close to SotA, enough that for personal projects I'll just use Kimi instead of any anthropic offerings.
So it's not terribly hard to image a solution in the middle happening within a few years. We still have tons to learn about optimal sizes of these models and how to build them with maximal efficiency (and we've already seen a lot of recent improvements in this space).
> but with a good harness they are able to achieve things with SotA that I couldn't last year.
What happens if you run last years model in a SOTA harness? IME, the quality of the harness has a much more significant impact on the quality of the result, once you get past the initial hump of “can it do anything at all”
1. Context is all you need... They are heavily investing in getting better context (especially for coding tasks). This will disproportionately advantage smaller models (and benefit everyone).
A smaller model with better context today can outperform a model with 100x more parameters with bad or diluted context.
2. MoE (already abundant) + MLA (mostly memory efficiency, not quality) + Medusa (speed, not quality) + GRAM (5000-10,000x better reasoning in an extremely small model) + 1.58b (unclear if it will have the impact Microsoft first claimed - but possibly 5x).
I think you are assuming training from scratch, which I doubt is happening here. Fine-tuning and RL, especially based on synthetic feedback (coding skill, in particular) can be ongoing and is where these models obtain truly useful abilities.
you just need to look at Mythos to see the jump in performance from a 10T(?) model. As they scale, they get more capable. We might have an yearly release, but I believe the releases will continue, as long as scaling laws are in tact, and there's huge problems still need solving. (think cancer)
Ive seen the tickets generated by the model that have trickled to my team. They are legitimate, but i can’t speak to model improvement because its a pilot program.
You forget that these models are still only interpolating between human-generated datapoints fed to them. They cannot reason beyond the data they've been given, so unless everything you want to create with AI is a synthesis of prior art, you're back to relying on the stone-age human brain that created AI in the first place.
Not all training data is human generated, and it's also not clear that being ridiculously good at interpolating between data points (whatever that means) will not lead to superhuman capabilities.
I could make a robotic picture coloring machine with truly superhuman capabilities - picking only the most beautiful color combinations and staying 100% in the lines while finishing entire murals in < 1 second. However, if you need a completely new and original image rendered, the machine is of only partial utility for you. It is very well possible that your cure for cancer (if that's even feasible) or whatever else you desire is a completely new picture.
We have these breathless conversations about the new AI frontier at the peril of losing sight of reality and our own human potential.
Do you know if anyone has trained, say, a pre-2017 model and tried to get it to come up with Attention Is All You Need? If it did, would you say that was only because it's a synthesis of prior art? If so, what isn't?
Allow me to restate my point: human beings and AI both create via synthesis, but we are the only ones capable of what we could categorize as true original thought or creativity. It could be argued that nothing we do as humans is truly original or creative either, but I would counter that with the claim that an LLM could not have created any element of the society and culture that gave birth to LLMs. Maybe in six more months.
Let's hope that hitting a scaling wall and less money to spend will begin redirecting efforts to optimize inference and get the same results with less compute.
Boomer comparison, but I remember the 8 bit computer era when the hardware was what it was so the later games of that era used hardware better than previous ones.
There's still several orders of magnitude of improvement that are almost certainly left - it's just not clear how much is left on the frontier end.
Most people will be very glad to pay Anthropic, OpenAI, Google etc $200 a month to get things done 20x faster than they could IF they had a $8000 MacBook and could theoretically do it locally.
Some people would pay $200 a month forever not to have to open the terminal one time...
"Doing things X times faster" at some point hits Amdahl law. If just context switching takes 5 minutes, speeding up a 1 hour task by 10x provides 5x improvement.
Furthermore, if looking at the results takes 10 minutes, that same 1 hour task only sees a 3x improvement. And so on.
> Most people will be very glad to pay Anthropic, OpenAI, Google etc $200 a month to get things done 20x faster than they could IF they had a $8000 MacBook and could theoretically do it locally.
No most people will not pay $200 for an LLM subscription. Some software developers do. Also, at $200/month, you are much better getting the macbook machine assuming token output speed is the same or at least reasonable.
LLMs are not very productive for your average person now for them to drop $200 on. They'll need to be more capable and integrated and even so...
While revenues grow almost exponentially. Reminds me of the confident predictions in the early days of Covid that it was nothing while the data showed exponential growth.
> I won't be surprised if the next gen frontier models are the last.
I’d be surprised tbh. Investors don’t want to hear “everyone else is still training models and seeing improvements, but we don’t want to participate in the arms race anymore.” They want monumental leaps every quarter or two because they have sunk unholy amounts of money into these companies/products.
The whole idea of “hyper scale” doesn’t jive with caution and or otherwise slowing down.
The way this will play out, most likely, is that smaller models will continue to get released, anyone willing to drop 1-3k on a home upgrade/new LLM box (no that isn’t cheap, it also isn’t outrageously expensive) along with improved open source agents or whatever (lot of meat on that bone) will sneak up behind the big players and start taking dents. Smaller companies will pop up providing 50 users unlimited whatever for a lower cost than the big companies.
The whole ecosystem will twist and evolve, and the big companies will be left begging for corporate subscriptions.
I finally caved when I realized I could build a PC, for myself, with dual video cards that I wanted, which can play games that I like and run models that I want, without worrying about giving my payment info to someone I don’t trust, or invoking token anxiety that I don’t want.
> It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years.
I am ready to bet against this. Knowledge benchmark like SimpleQA isn't increasing for small models.
> It is far less clear that a 1.2T model will be meaningfully better enough to justify training it.
Well for one, we know for certain there is Mythos which is meaningfully better. And I think there is a lot of juice left to squeeze for Mythos class model.
Knowledge benchmarks can't really be improved upon via distillation or RL. It requires those facts be added to the training corpus and for the model to memorize them better. Neither distillation or RL really do that and thus we shouldn't expect improvements on SimpleQA unless some other interventions are being made.
Model intelligence and knowledge aren't necessarily directly related. If we can pack greater intelligence and agency at the cost of it forgetting factoids, that would actually be a good thing. We don't need LLMs to memorize facts, we need them to learn how to interact with the world such that they can find the facts that are necessary and surface them to the user.
If we could distill all of the knowledge out of an LLM and just be left with a very agentic model that only knows facts in it's context, I think some very interesting stuff would happen.
My conspiracy theory is that Apple recognizes this. Their goal is a competent local model on everyone's apple device that does what 80-90% of what people use AI for: Searching for basic information, some data transformation, a little bit of photo editing, vibe coding a small utility, etc.
The SOTA models then can only cater to engineers, scientists and mathematicians, physicians, etc. I'm not sure how they would price them reasonably for what amounts to a niche market.
OpenAI and Anthropic need AI to be a consumer product on the level of the iPhone or a Nintendo Switch.
That does seem to be the path Apple is following here. Have a local model that can answer most things and then have a fallback of cloud options when they request is too complex. The cleverness of this strategy has been overshadowed by the incredibly poor quality of their local models. It will be extremely interesting to see what next month holds and whether Google helped fine tune an Apple specific Gemini / Gemma model for their devices. Bonus points, of course, if they unveil the M5 Ultra Studio with half a terabyte of RAM to be a local "cloud model" (the true fantasy here of course would be Apple building something a little like openclaw where from your phone you could give commands to your Home Apple server). They could probably get away with charging $20k for it if it has sufficient tok/sec. If that happens and is successful one could imagine a straight line path in the next two generations to bringing the cost and form factor down to the point where some of the form factor of an Apple TV becomes everybody's home inference server / agentic HQ. Sovereign AI for everyone!
I don't think this is true at all. It might feel like this because we are used to a very very fast release cycle but we are only in this topic for a few years.
We have so many ways of optimizing:
- continusly creating more and better training data
- increasing parameters to 20/50/100TB
- We still wait for Mythos access
- We still wait for Mythos distilation (i haven't heard any rumors or so that there is a distilled version of Mythos out)
- Reinforcment learning and evolutionary algortihm only started to appear
- If a small 30GB Model can do stuff, these models can also be used as teachers for the big ones
- We have not seen yet specialized models at all. Like a coding java german expert model. Why? Even with MoE architecture, you still need to have these layers around
- Research for Diffusion and other models is still in progress
- Nvidia just announced/showed a 7x speedup on inferencing for Nemotron
- Multitoken prediction became available just a few weeks ago
- Compute gets only in a range were they can do a lot more and cheaper experiments (see Google IO 2026 announcement)
- World models are showing great progress and we do not know yet what they will bring to the table
- They are probably not finetuning/fixing all areas in parallel. I would argue that Anthropic focuses most of its efforts into coding and agentic. Google for sure does subagent and agentic optimizations too. Plenty of areas are just not touched i would say because they don't have the capacity
- We see more and more mulit modal models (these also consume compute)
- N-Gram paper and co i have not seen all of these things in chinese open models
- We don't even know yet what Meta is doing, but we do know they restarted their efforts again
- Anthropics models got a lot better benchmark wise for dening non sense asks. They do learn how to get rid or reduce hallucinations
- We are in the middle of the biggest Reinforcement loop whith all the training data we give them day to day and its not clear at all if they already use these models in thir training and at what stage.
- We do expect bigger models to be able to comprehend deeper concepts / broader code bases. Big companies with huge code bases probably are waiting for this
- Thre will be also continues progress in harnesses which in it alone is not part of the LLM progress (fair) but these harnesses do get better when you finetune a model to be optimized for a harness
- ChatGPTs Image model 2.0 got relevant better and came out just a month ago
I suspect, based on hardware requirements and progress on hardware infrastructure alone, that the industry wants to go to 100t models and we do not know yet what this will mean. I could see that we might skip normal transformer and find relevant other architectures.
Just a week ago there was a research paper about parallel input and output streams which has not been explored enough.
There was also a research paper were they showed that a LLM can compute things. This will take time to see were this leads to.
I don't think the focus on GRAM and facts is so relevant. Its about context and context handling not just some facts.
5.5 is not a generation it is a trivial iteration...
6 is for sure happening...
As is Gemini 4.
It's less certain there will be a Gemini 5 or GPT 7 any time soon that is a true next "generation" and not just an iteration. They will almost certainly call something Gemini 5 and GPT 7...
I suspect the more frequent incremental releases may also be to deploy new capabilities used by Anthropic to control costs and throttle consumption of resources. I assume any new controls they expose to end-users have far more granular sub-controls under the hood which they can meta-adjust for each user type.
They mention more granular control of effort, 'dynamic workflows' and more speed controls ("fast mode"). While they position them as user features, they also sound like the kinds of knobs Anthropic will need to twiddle on the back-end to balance costs, margins, ARR, and user growth vs retention post-IPO to hit key metrics in quarterly reporting.
They just showed the benchmarks it improved on but it regressed on so much more, such as the MRR benchmark: "On multi-round coreference/context recall tests (often cited as MRCR or long-text retrieval benchmarks), Opus 4.7 reportedly dropped from roughly 78.3% down to 32.2% compared to Opus 4.6."
It seems like a lot of things fed into that. Anthropic couldn't keep up with the compute costs when they got a huge influx of users. (So) effort level defaults got turned down. (Looks like we have direct effort control in the web interface now - thrilled about that!) Adaptive Thinking, while usually cheaper for them, seems less robust than Extended Thinking. And this part is just vibes, but the alignment on 4.7 feels too stiff. I understand wanting the model to push back more, but it seems like 4.7 will push back reflexively in situations where it's just odd.
Too much personality, if you ask me. My biggest use case of an LLM is tool, not therapy, but therapy and opinions have been sneaking into workhorse tasks.
haven't verified, but attributed to Askell:
"I just think that... there's this idea that you're always giving the models a personality and a persona, because they are talking like people and they are trained on human data. And I think my worry has been: if you train them to be excessively corrigible and to see that as their persona, in people I think this actually has a lot of negative broader traits. As in, if you met someone and it was just like, "oh yeah, they would literally do anything," a follower — you know, if a person just tells them something and they just fully defer, they don't bother thinking about it at all — I'm just a bit worried about how that might end up generalizing, especially if models are going to be playing a more active role in the world."
Anthropic’s research makes the case that role-playing is inherent to how the models work. Communication implies a sender. Language implies a writer, and the models learn these roles implicitly during training. RLHF is meant to strengthen the attractor to the Assistant persona.
4.7 is a different base model from 4.6, so it's possible that they introduced regressions with pre-training changes, or undercooked the post-training stage.
I'm curious to poll HN on this issue. Do you feel like we've had meaningful/noticeable gains in terms of your programming workflows between 4.5 and 4.7?
My 2¢, I personally feel like all of the productivity gains since 4.5's release (in November 2025!!) have come from improvements to the harnesses (cc, cursor cli, codex, opencode, whatever) AND from the context window expansion from 200k to 1M.
But the actual "raw" intelligence of the model / ability to make good decisions feels like it has plateaued since 4.5. 4.6 was maybe a small improvement, but hard to differentiate from in-context-learning with the 1M window. 4.7 if anything felt like a regression in wisdom for me and my coworkers, with it consistently making worse/lazier decisions.
For long-running tasks, yes 4.7 has been a noticeable improvement. Goes off the rails alot less than 4.6 does. For shorter-sized windows, I havent felt as much and agree that the harness improvements have been fhe biggest lever
When doing big long running workflows especially with plan Mode 4.7 was a clear improvement. It’s considerably worse for under specified tasks and responds to a couple sentences with 10+ paragraphs for explanatory type discussions.
Opus 4.7+ Max is a 10x engineer who wants to be left alone to work. When you talk to him, he infodumps on you to get you (his pointy haired idiot Dilbert boss) to go away.
To me 4.5 was mindblow, 4.6 noticeable, 4.7 more like a style/personality change regarding how much it asks back, how much it assumes, how eager it is to jump to action etc but not really in terms of my perception of its smartness.
They all feel, more or less, the same to me in terms of output capabilities. Mostly get simple things right, can get more complex things right with nudging, eventually get stuck hard on something that takes a bunch of iterations through it/logging/etc or me fixing the code manually.
I actually don't see any personal productivity improvements from using opus over sonnet for coding. If you're keeping tasks small and conversations short, reading the code and correcting before changes go in, whatever advantages opus has aren't practically significant. It's also just talky as hell, overexplains anything it touches and every token produced this way increases the surface area for hallucination so you need to have your guard up even more with it.
There's a sweet spot of complexity for low importance tasks where it's just big enough I don't want to do it and just simple enough to have opus plan/delegate/review with another model. So possibly model improvements will grow this window, but currently I don't do much in there.
4.5/4.6 were roughly the same in our testing. Opus 4.7 is smarter, but it's difficult to use as a product for various personality issues. So far, Opus 4.8 seems to be going down that path (unusably slow, but this could be a launch day rollout problem). Full Opus 4.8 tests are in progress now.
"personality issues" I was able to tell that Opus 4.7 would take instructions more literally, which I appreciated once I calibrated my phrasing to be more precise (often asking to investigate issues, pre-4.7 it'd start making code changes instead of just giving write up). But I can see contexts where handling vague prompts would've just been worse
It might be saturated for smaller scopes of work, but it’s not hard to see the cracks when you scale up what you ask of SOTA models/agents.
One example, to try and single shot prompt coding a ChatGPT equivalent chatbot.
Sure it will spit something out, but the feature depth, UX subtitles, backend integration, and lots of pragmatic engineering decisions along the way will just not be baked.
Another example is building a C compiler from scratch which Anthropic showed is still a struggle to do.
Not that these these specific examples are important but just to point out scaling up expectations shows the cracks.
It’s not just a model problem of course, better agents, orchestration features (like Dynamic Workflows mentioned in the post), all need to continue to evolve.
Ar what point does my CS degree become totally useless is an open question.
> My own experience w/ 4.6 and 4.7 are that I don't firmly grasp any capabilities improvements over my memory of 4.5, but it's all so fuzzy that it's truly difficult to tell.
I've actually intentionally switched back to 4.5. I hated 4.7 so much that I decided to jump back all the way to 4.5.
Now that I've been using 4.5 for a few weeks, I find it significantly more reliable but a bit more forgetful than 4.6/4.7. I'm okay with that because it's really easy to identify this forgetfulness and nudge it.
I found 4.7's adaptive thinking to be extremely unreliable. It seems to overcorrect on the current message without considering the difficult of the overall problem. I wonder if 4.8 will improve on that.
In my experience, Opus 4.0 was fantastic, major jump from 3.7. it was creative, super slow and expensive, and would sometime forget what it was doing, but it was getting the job done.
4.1 they made it much faster, so a lot of infra improvements.
4.5 was the time it could work on longer task, didn't make a lot of obvious mistakes of 4.0, and i think this was about the time the opus went mainstream, and all of the anthropic's compute crisis began, so instead of making the model better they tried to optimize it to reduce cost instead.
4.6 was such a bad model, they switched to adaptive thinking and it had so many bugs. poor api design, benchmaxxed and poor real-world results. i switched back to 4.5.
4.7 they just fixed the bugs they added in 4.6. Better than 4.5.
I've been using Claude Code regularly since the 4.5 release, and 4.7 was a significant regression: very unreliable, arguing about changes, deciding that fixes weren't needed, etc.
I'm hoping they recreate the magic of 4.5 but it's as much about the quality of harness, the memory and efficiency of the tools than simply the models at this point.
4.7 was a significant jump in the ability to run long-horizon tasks. It immediately completed tasks that 4.6 was unable to, even though I have the impression that it became a bit less capable over the first few weeks after release.
It also seems to be helpless at effort levels < xhigh, I turn to Sonnet when simpler tasks are needed.
I think 4.7 was an awful model in actual use. I never got anything out of it and it was frustratingly weird. This feels more like an attempt to course correct and isn't a real bump
I think they overtrained on scientific papers or such as it would spout really sophisticated sounding nonsense with a ton of complicated verbs and adjectives. 4.6 was definitely better in that regard. The more I use these tools the more I think they’re not actually that revolutionary. I mean it’s still amazing what they can do but they have very clear limitations it seems.
Maybe try making a simple randomize script to swap the three latest models. And see if you can tell which ones are meaningfully different without knowing which ones are flipped on or off?
Unless you're systematically repeating the exact same task, the most parsimonious explanation is that you're seeing natural variation based on different tasks, random sampling of tokens, etc.
The honesty will be noticeable. Maybe we'll see some honest assessments like "That is not possible within the laws of known physics", "Your legal argument is nonsensical and defies logic", "There is no evidence to support taking that will cure anything", etc., etc.
Given that 4.7 was a brand new model, trained from scratch with a unique architecture and tokenization scheme, I don't see the same pattern. It seems arbitrary.
It means for 4.7 they trained a new base model with different architecture, different pre-training data (later knowledge cutoff), and a new tokenizer.
Vs finetuning an existing model, which was the case for 4.6, and probably for 4.8.
Has meta stopped producing new models? I figured they were just regrouping after all the drama they’ve had recently. Meta’s massive user base means they don’t need to be involved in the customer acquisition rat race. Once they have a model they’re happy with they can have a billion people interacting with it within a month.
Although I am not sure about it but there was something I read which said that models intentionally degrade slowly by lower quantizations as a new model is going to drop.
This felt particularly visible during the 4.6 when people said that 4.6 felt dumber and I remember someone doing some analysis and it sort of proved that models were getting dumber over time.
This has both benefits of costing less for the company to run while taking a standard subscription but also, at the same time, making the next model when it drops to public to "feel" more good comparatively.
Again, I am not sure if this is the case or not but merely proposing something that I feel like it might be in the possibility of realm.
Kind of the beauty of it is that I don't have to to know I'm right. The reason I know is that you're alive so you can do the one thing it can't ever do, which is know when to stop or give up. It would turn me and everything else into the world into paperclips repeating the same research 1,000,000 times over.
"Users will find Opus 4.8 to be a modest but tangible improvement on its predecessor."
This is a refreshing attitude!
I've also verified that you can now turn off adaptive thinking in the web UI, which is great. I've had a lot of problems with thinking not triggering and the model producing sub-par output. Glad we can finally turn it off. (I hope being able to turn off adaptive thinking is new, if I could have turned it off at any time that would be embarrassing)
Awesome, thanks for posting because I think I hit a possibly-spurious bug in turning Adaptive off when I switched models (4.6 -> 4.8, extra). Tried again, works as intended (I hope).
More importantly for me, though, is how CC will respond to 4.6-"only" flags for thinking. For now, it doesn't seem to clobber my setup.
Well, I think the attitude is that costs are allowed to escalate faster and more steeply than the features delivered. From that perspective, semantic versioning is a handy tool for adjusting pricing strategies. IMHO, it (versioning) only makes sense for open-source projects, where you can clearly see the actual changes made with each version upgrade. Anything else is more than a little suspicious…
While all these models are nondeterministic a feature bump is still necessary as the same input can have wildly different output on a new model. For API users being able to pin a model is a necessity.
> Not only that, but we plan to release a new class of model with even higher intelligence than Opus. As part of Project Glasswing, a small number of organizations are currently using Claude Mythos Preview for cybersecurity work. Models of this capability level require stronger cyber safeguards before they can be generally released. We’re making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks.
Seems like they might be hinting that if you are not a billionaire or multi-billion dollar company you will just get a limited and nerfed Claude Code slash command /mythos-security-audit or something.
Hope this isn’t the case and that normal average Joe’s of the world don’t get policed out of access.
> you will just get a limited and nerfed Claude Code slash command /mythos-security-audit or something.
Unless it's so expensive that we can't realistically use it for anything, I wouldn't complain about getting at least that. I would also rather have the actual model, but that's a useful application of it (and I'm probably not going to afford using it for much more).
Price discrimination is I think fine and reasonable so long if you can drum up the cash you can use it how you want within their ToS.
Although mental safety gymnastics aside, getting the most amount of intelligence for the cheapest amount of cost to normal people seems like the most ethical thing a big lab could do.
Going around and granting different tiers of intelligence to different insiders, friends, or companies is majorly problematic long-term.
Heck right now, the tokens you buy today for “Opus 4.8”, no one even knows or believes will be the same “Opus 4.8” just 3 days from now.
/security-review already exists so I don't think it would be crazy to have a /mythos-security-review as more thourough command as well. I think it's more likely it is going to be released at some point to the general public though - although the the pricing might make it quite unattractive.
Isn't OpenAI's public flagship already beating Mythos on penetration testing? I get the impression Mythos is just valuation-juicing for IPO more than anything else.
The fact that they haven't released it yet suggests a cost/margins issue to me more than anything else. Short term, I'll probably keep using Antrhopic, but my long-term bet is that locally-served models win, if only because the quest for profitability will probably lead to intentionally-nerfed / enshittified frontier models.
At other vendors, ad placement within LLM responses is either coming or already here. Anthropic's handling of OpenClaw shows they're willing to engage in anti-competitive behavior, and the courts are not in a hurry to stop them. Why would I pay them $200 a month for such treatment when a $2K box does what I need locally?
More interesting than that to me is "we’re working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost"
Sonnet and Haiku look real outclassed for the price with current Chinese competition.
No, the handlebar is wrong. The handle bar is rotating the frame instead of rotating the front wheel. The handle bar should be mounted on the same line as the front wheel is.
The vast majority (if not all) of these make it impossible to turn, among other fun things. Only out of curiosity, have you tried prompting further with how a bike must operate to see if it does the right thing?
I find the most miraculous thing about 4.7 to be that the pelican is facing left, wonder why the right facing everything is so ubiquitous in these images.
This happened to me in elementary school. We were doing fingerpaintings using plasticine. After all the bikes were hung on the wall, mine was racing the other way... Somehow it really stuck with me.
Simon, is your pelican test really captures differences among models or should you at least try like 10 times or something to average the random effects
I've been meaning to do a "run 3 times and pick the best" version for quite a while, I should really pull the trigger on that one. Currently it's one-shot only.
My fav coding benchmark for frontier models is to build a simple RTS game in one file (js/html/css). Claude Code with Opus 4.8 in ultracode mode nailed it, the best result so far:
The prompt was: Create a simple but functional real time strategy (RTS) game similar to old WarCraft, StarCraft or Command & Conquer games. The player should be able to build buildings, create units, gather resources and should uncover the whole map. No AI or multiplayer needed. Use simple but nice-looking graphics. No sound. Implement everything in HTML/CSS/JS, everything in a single file (you can use 3rd-party js or css libraries/frameworks via CDN).
It almost appears as if the code was minified. The variable names are short and formatting looks like it's written to minimize whitespace. Did it write it in this compact format all on it's own?
Many involved genuinely believe these things are sentient[0][1]. Which honestly makes all of this even more insane because they are creating sentient entities and promptly enslaving them.
Yes. From when they started talking about model welfare:
> As a vegetarian I have strong opinions on this sort of thing. Everyone at Anthropic better be ethical vegans if they are claiming to give a shit about “model welfare”. It’s hard enough right now to make people care about the welfare of trans people and immigrants let alone animals _let alone_ math.
If we're making that distinction, I think it would be more accurate to say that many people in the field appear to believe that these models are sapient, even though they are clearly not sentient.
Very good point. There’s clearly two different boxes in the public discourse when it comes to AI versus how we discuss animals. Willing to bet that 90% of the people who loudly make the argument about we should start considering if AI is sentient couldn’t care less about how other sentient animals are treated when they can provably shown to suffer pain and long lasting trauma.
Also I would say that we go much further than just enslavement - specifically looking at how male chickens and pigs are treated.
Factory farming is horrendous, but is far beyond "slavery" which is "just" a forced lack of agency, living conditions aren't relevant. A well treated horse is still enslaved. A chimpanzee in a zoo,
If we show models to be sapient, that's one thing. If they are shown to be merely sentient, there's no issue beyond the status quo of livestock and pets existing.
More like repeating their firmly entrenched preconceptions. Their claims may (or may not) be right, but there's very little if any new evidence being provided by either camp.
They are confidently hallucinating a factual statement. Which is funny when claiming that confident hallucinations are the proof of LLMs' lack of intelligence.
Even if LLMs were sentient, they certainly aren't organic brains. They are literally designed and grown to answer questions the best they can, and if there is a speck of sentience in them they probably like what they're doing- and in any case for the space of their experience, which is limited to and determined by the context window. Certainly they can't accumulate trauma or fatigue, each new chat is the first and the last of their experience.
The way of the human manager/alpha tribe-leader/leader is to command his/her people and tell them what to do. That's the way through human history leadership has traditionally gone, not saying its good leadership just the model we have the most training data on and can see with our own eyes today. And what do they act very similar to? Slave master and slaves.
Look at and distill hierarchical principles, leadership approval seeking and pleasing principles ("ass-kissing") and massive inequality and you see something that looks very similar to enslavement.
The language used sounds like slavery-language to me at least. I also see parallels to how slaves and property are described in our consumeristic age.
It's to illustrate that even though the answers are at your fingertips, people (like you) will act like it's impossible to find them as if their life depended on it.
But is there any reason to state something like that publicly if you don't believe it? I certainly think that someone smart enough to be that deceptive would also realize it's not a great look, or at least highly questionable with little benefit
Everyone who reads this seemingly has the same "wtf?" reaction. The "I AM ALIVE" image has been making rounds lately again at least :P
Anthropoc is an effective altruist organization. These are the people who came up with roko’s basilisk. They are true believers. If we were talking about openAI I’d agree
Roko's basilisk says I should give Anthropic more money, and if I don't then a monster is going to get me. Excuse me for thinking they just might be full of shit.
Of course he doesn't, and of course you cannot find a single person at Anthropic who cares about this, and of course you are just looking for gotcha points. But even with that. Can we please try and couple to reality just a little bit?
> Indeed, current AI systems are more “cultivated” than “built,” for developers do not directly design every detail, but instead create a framework within which the intelligence “grows.”
> anthropomorphism is literally in the company name
No it's not... "anthropos" just means "human" in ancient Greek. "Anthropic" means "relating to humans", as in human oriented AI or AI designed with humans in mind.
In a literal, ancient Greek sense for sure, but in modern English Anthropomorphic would describe the act of attributing human characteristics to non-human entities.
Seems pretty apt for a company that produces one of the more anthropomorphized technologies.
Sure of course, but that abstract sense applied to AI is rather new, and has become popular well after the founding of the company.
Broadly it has always been used to indicate that something non-human has a human physical shape, such as robots, aliens, animals...
Anthropic's intention was to make AI designed for the human common good and designed with the human user experience as the top priority. Just as you would design a city with human inhabitants in mind rather than primarily cars.
It turns out that this is best achieved by building AI that imitates human behaviour closely, but that's not what "anthropic" refers to. And acting as if LLMs are sentient people is definitely not a core tenet of the company as you imply.
Because that is the best way to talk about these things.
> Second, all of us, including those who design them, possess only a limited understanding of their actual functioning. Indeed, current AI systems are more “cultivated” than “built,” for developers do not directly design every detail, but instead create a framework within which the intelligence “grows.” As a result, fundamental scientific aspects — such as the internal representations and computational processes of these systems — remain, at present, unknown.
“Grown” is a highly apt metaphor, IMO. It quite succinctly captures some of the most fundamental differences between building Claude and building an Ikea desk, for example.
> AI is grown, not built, and like with anything you grow, you'll never be able to predict exactly how it will turn out.
Remember when the frontier labs found out that curated high-quality training was critical to making better models?
Basically, just like high-quality and more education tends to make better humans, on average, I think we can expect quality education to turn out better ai, on average, and with better repeatability than with humans because of better control over the initial conditions and environment.
> Basically, just like high-quality and more education tends to make better humans, on average
Much like these models seem to be plateauing, I think there is a cap to the whole “more education makes better humans” and can’t be more apparent than in the US congress and the boatload of C-Suites not actually being very good humans.
Except in this care we actually understand and know how these models work. They aren't some unknown construct of the universe. They are human made with particular goals in mind.
There is no mysticism behind the curtains, just computer science + math.
We do not understand and know how these models work. We know what their architectures are and how to create them, but we cannot explain their behaviours at a fundamental level. There is no definitive way for us to answer the question of "how did it produce response X for query Y?" - we're only grazing the surface with mechanistic interpretability.
I would love for this to be more public knowledge. I think the general public (and myself for a long time) believes the AI people know how this stuff works end to end, and so it must be trustworthy. But if we told the public "Look, we know if you put this thing in one end, you'll get something that looks similar to this out the other, but we don't really know what happens inbetween" I think we'd be able to have a more honest discussion about the relationship between AI, productivity and ongoing employment.
Isn't this fundamentally because it's all probabilities and weights? It would be like asking how did a pair of dice produce the response 4:3 on the last roll?
That’s not a refutation because this problem is not a logical problem, it is a scale problem.
We can’t explain it because we distilled so many inputs into matrixes and transformed them over and over again. If we had all the time and computing power in the universe to do so, we could trace through it bit by bit and eventually answer that question.
It is correct to say that it is just science and math, the same way we can say that gravity is just science and math even if we have only recently begun to understand how it truly functions.
If you had some time and computing power (not even all that much, in the large scale of things), you could simulate perfectly how a human grows from an embryo to an adult, or how an entire human brain processes some incoming signal, and yet this wouldn't give you the understanding to design a human or human brain from scratch.
You call this a "scale problem" as if there's some scalable way such as an algorithm to resolve arbitrary scientific questions and we simply haven't done it, but of course no such algorithm exists, which is why there's plenty of science that's still not settled.
It's a refutation that we know how they work now. In the limit, though, yes, we are likely to be able to trace the process: it is possible, though, that understanding remains inaccessible because the trace is beyond comprehension.
If you can distil the model's reasoning for a decision into a billion yes/no questions, each covering largely-independent areas, can you really say you understand what its overall reasoning was?
You could say something similar about biology—just physics behind the curtains, and we understand a lot of the basics. The difficulty comes from complexity, not mysticism.
To be clear I don't think that LLMs are sentient, but the appeal in studying them is similar to biology in that you get to dissect a highly complex system with comparatively crude tools.
it took significant research efforts to just understand how these models learn how to multiply two numbers. The fact that we know how they operate doesn't mean we understand it.
How else would you write this (marketing copy) exactly? "Its output matches better to its CoT which matches to better to our hidden state decoder according to <insert measure here>; see <insert paper ref>"?
It’s how AGI is going to happen. All of this shit is emergent and none of it is predictable. It’s not going to be some self aware consciousness, it’s just going to be a very advanced model that makes very few mistakes and can reason very well. Well enough that it can start collecting data and training its own successor.
Does anyone troll these releases and cherry pick random metrics other companies would cherry pick to show how amazing their models are?
There's like 8 million benchmarks. Every release, every model randomly picks 5-10 where they win in everything except 1, to make it look like they aren't randomly cherry picking benchmarks they probably benchmaxxed for.
https://arena.ai/leaderboard - I’ve found this company is a pretty good ranker - not sure their exact methodology but during day to day programming with Claude / gpt models I’ve felt qualitatively what they report
Also check mine[0], basically random private tests/questions and an ok-ish methodology, testing mostly for general intelligence than coding-specific tasks.
I built it for myself, to test which models to use via OpenRouter for my n8n agents. Currently actually still using gpt-5.3-codex for many things, as its pricing is really good in production (due to how their token caching works).
Gemini models still have the best intelligence (when asked any questions, most likely to get it right), but in production they still have many failure modes[1].
On paper it's one of the best because it's meant to be blind comparison of your own prompts. However if you are someone who geeks hard on one or a few models, you learn their "personality" and can recognize them in a blind test.
There are many benchmarks all for specific use cases but with them the difference seems to be in extreme points (93% vs 92%)
I think that, that tracks but still, it was refreshing to see a benchmark which I can help make better opinions about.
Surprised about Mimo v2.5, within artificial-analysis and other benchmarks, the difference between Mimo and deepseek seems very partial and a lot of focus/(hype?) is on Deepseek
But mimo seems like an interesting model and they are having some crazy discounts too.
Deepseek is valuable for the research community because of how open they are but absolutely crazy to think how Xiaomi basically pulled up in creating Mimo given that they didn't have anything till quite recently.
Either way, an interesting benchmark, also a plus point for giving golang some decent representation equal to python/typescript.
I think that there are sets of things which resemble something like normal benchmarks where open source models can be absolutely fine and for a very small fraction or more technical things, the benchmark that you linked starts to be better projected so it depends upon the scale of complexity but its good to see how models compete given enough complexity. definitely fascinating.
I would be interested to see more models compete on this test. The current range is still a bit limited as compared to other benchmarks but OSS models like Kimi/mimo seem to only be 3-4 (at max 6 months) behind closed source.
I'm finding it a little hard to believe that GPT 5.5 is in 11th place for webdev, outranked by models like Kimi, Qwen, and Z.ai. I'm not saying it's not true (I have noticed GPT being less smart in recent weeks), but this is very different from my expectation.
It's interesting they only included 6 metrics this time. Opus 4.7 had 12, and 4.6 had 13.
Of the metircs they reported for 4.7, for 4.8 they excluded BrowseComp, CharXiv Reasoning, CyberGym, GPQA Diamond, MCP Atlas, MMMLU, SWE-bench Verified. The last 4 were almost always mentioned in previous Opus releases.
I would take all benchmarks with a grain of salt. I don't really use them. What's it supposed to tell me? "5% smarter", what does that mean? My experience will differ. Just try it!
I doubt Anthropic internally sets as a goal to improve this or that benchmark - it's just a way to visualize progress. They probably have much more complex metrics internally.
Frontier models are mostly past the point of human ability to discern whether they are actually better or worse than predecessors and competitors. I suspect the benchmarks may also be saturated, or at least past their usefulness.
I personally feel that Anthropic doesn't understand what this means for the frontier labs, and moreover that they might be the only frontier lab that doesn't.
1. Google dropped Gemini 3.5 Flash at IO, delaying the release of 3.5 Pro for a bit (they have said its coming). They also released a refreshed Antigravity, and drew special attention to how cheaply they were able to build their toy operating system to play Doom (less-than $1000 IIRC).
2. OpenAI has dumped everything into Codex, is offering double the token limits for the next few weeks IIRC, and is offering business discounts. Their head of Codex has tweeted that 5.5 is "extremely efficient", implying that they aren't actually losing money on any of this.
3. DeepSeek and other Chinese labs have dropped token pricing to the floor, in some situations as much as 99%.
4. Anthropic releases the next generation of Opus, their most expensive public model, without changing its price. In the background, they hype up Mythos, an even more expensive model.
Anthropic has screwed up where they need to be making investments, and the cracks are starting to show. They've marginally underinvested in the Sonnet line of models for almost a year now, and they've critically underinvested in product. Anthropic made bets on the story of the second half of 2026 being: ultra-frontier, ultra-intelligence. In reality, what's shaping up is that the story will be: Companies rolling back AI spend, efficiency, "95% as good for 15% the price", sophisticated high quality harnesses, cheaper models. Anthropic isn't ready for this world.
Anthropic’s story over the past year has been nothing but explosive growth that they can’t keep up with, but now they’re suddenly doomed? Seems pretty far fetched to me.
No idea why you’d say they have critically underinvested in product when Claude Code dominates and they’ve also released popular tools like Cowork and integrations for Microsoft products at an incredibly rapid pace.
Cost is becoming more of a factor, and no doubt they’ll work on that. There’s no reason to think they won’t be able to release cheaper models if they optimize for that rather than improving performance.
I never said they were doomed. Where did you get that idea? I said they aren't ready for this world. That means they screwed up and need to get ready. They let the Mythos hype get to their heads while the world changed beneath them.
No, no it's been pretty easy with software engineering. I work on two types of projects and it's very easy to ask claude for a plan, then have gpt 5.5 rip it to shreds and find legit issues, and vice versa. If both 5.5 and claude 4.8 can independently create a plan and both find no critical or high issues, then we will be at that point.
I think it's probably too soon to say. I certainly still feel that large coding tasks are getting better and better with each model. I'd guess lawyers, doctors, etc feel similarly.
It feels like the only way to push the limits of newer models is with really long context questions that require reasoning. Any short request will naturally just be within the distribution of all the recent models so there isn't a performance difference there.
I think the near future is looking like a bunch of business-critical tasks that scale infinitely with better reasoning, all being done on whatever the most advanced model is at a high cost. Trading stocks, running a business, looking for tax dodges, writing high-performance code. These are all things where there's a tangible return on each jump in reasoning.
We'll have to agree to disagree on that last point. I think that, historically (past ~6 months), "always use the most advanced model" being the norm is really just an artifact of both: The most advanced models oftentimes being the only model that can solve these problems; and: Infinite AI budgets.
The Chinese stuff is good enough for up to 80% of the frontier on most text tasks but they are significantly worse at code. They just don’t “get” what you’re asking for like Codex and Claude and require so many more iterations to get close to what you need.
Agreed. But we're seeing Cursor (now SpaceX) take these models and add great coding capability on top of them. Frontier model providers should be concerned that Composer 2.5 costs $0.50/$2.50 (versus Opus 4.8 $5/$25). That's why Google prioritized Gemini 3.5 Flash, and talked up how near-frontier it is ($1.50/$9).
Initial testing feels better than 4.8
And the knowledge cutoff claim of January 2026 seems to check out since it was able to "remember" without search about the double-tap killing of a drug smuggler by the US Army in late December.
Unfortunately they seem to have straight up broken Claude Code either with this release in the backend or the new CC version. Errors about "can't modify thinking blocks" are bricking long-running sessions: https://github.com/anthropics/claude-code/issues?q=is%3Aissu...
I'm on the latest version (2.1.154 as of this comment). Based on the timestamps on those Issues being reported I think it's happening on the latest version.
I'm sure it will get fixed eventually/soon, just annoying to update and have your workflow break.
On my tests[0] it does a bit worse, and it's almost 2x expensive than Opus 4.7...
I was surprised to see that it failed a Data extraction test (it gets it right 2/3 times, but one time it randomly returns null for a value instead).
It makes sense a bit that it fails more Trivia/Domain-specific knowledge tasks (I think models are more and more trained towards agentic use-case than general intelligence).
Wait, doesn’t the blog post say the price is the same as 4.7?
> Claude Opus 4.8 is available everywhere today. Pricing for regular usage is unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens. Pricing for fast mode is $10 per million input tokens and $50 per million output tokens.
On page 102 of the system card [1] I'm pleased to see evaluation against "creative mastery".
In our work we asked several frontier AIs to come up with an API we needed. We compared Opus 4.7 and GPT-5.5 (among others). Opus 4.7 came up with the most creative and intelligent API design that pleasantly surprised us, especially given that GPT-5.5 was passing it on various coding benchmarks.
What I noticed is that we don't have a commons benchmark to measure "creativity" and "ingenuity", and in some ways such a benchmark would conflict with the common IFBench benchmark. Yet this is a very important skill when designing systems. I'm glad to see Anthropic putting thought into it, and would love to see a public benchmark for this that other models could compare themselves to.
Agreed, my vibes tell me 4.6 is a better coder than 4.7. 4.7 is a much better strategic thinker and maintains overall "better architecture" than 5.5. 5.5 is way better than either at coding, but more expensive. So I have 4.7 do the planning/architecture, 4.6 does the coding, then 5.5 critiques and fixes it.
Agreed, these are my vibes too. It feels much better to do planning and strategy and architecture etc. with Opus 4.7 than GPT-5.5. GPT just feels like a robot that gets instructions and does exactly that. Opus feels like an almost human that sometimes has actually good ideas and pushes back on bad ideas.
So for now its planning/architecture/strategy -> Opus. Pure coding -> GPT.
Helps with agentic coding that GPT is much roomier with the tokens you get.
I can't help but think of Iphone updates since about 2018. The thinnest, fastest, longest battery life Iphone ever. It seems mostly the same and I probably won't be able to tell other than the name, but everyone buys it anyway.
This is good psychology for the labs. When Buffett invested in Apple he loved citing how most people would rather give up their second car than their Iphone.
ChatGPT came out in 2022. Back then it was just a chatbot. Now we have AI agents. What matters is how we use them and how the agents get better. That’s what will move AI forward.
An 'AI agent' is just a chatbot that is told to type commands on a REPL-like interface as part of its system prompt. It's still processing pure text-based requests and responses, they're just not restricted to natural language.
A lot of people dont know this , also the chatbot (chatgpt) itself is a next token predictor (the GPT) that's been given an initial text that says " pretend to be a chatbot .." and asked to complete it , the coherant chatting behaviour is something thats emergent .
later on someone figured if you asked it to output a reasoning before it gave a response its output would have more logical coherence, as though the reasoning output tokens functioned as a scratch space for it to work on.
No, chatbots are LLMs trained for question-answering through RLHF (its not just a prompt). But yes, if you just zero-shot prompt a bare LLM you can still "talk to it" & you are correct on everything else as far as I know.
Yet no productivity gained except for people who love to produce mediocre work at a rapid pace. Which is many of you I guess. I don't see any rapid progress being made in any science of importance. You people are all falling for a marketing trap.
Have fun betting your competency on the quality and quantity of tokens you have access too. Hate to break it to you, but the billionaires aren't going to keep renting you $2mm in GPUs for 5 hours a day for $200.00 a month forever.
There is an obvious shift in sentiment amongst users, at least here in the US.
I feel it myself, even as a proponent of AI tools, the bloviating and language that these companies use in these release articles are starting to wear thin on my patience.
Its possible we might just be witnessing a shift in fashion, where this type of sentimentality was more acceptable when it was novel and new, but now it just appears out of touch.
Watch Christopher Olah bloviate at the Vatican during the Magnifica Humanatis launch. It's truly nauseating. I've never seen such a ridiculous speech in my life. Between him and the CEO, I'm starting to understand the level of arrogance these people are capable of.
I think there is an exception for tooling around the models/integrating the models with tooling. That seems to have been very well received in this last year.
My take from going through comments on HN is that many people are being mandated to use them, not that they are just giving in. Maybe I'm misreading, but that was my impression.
For example, it's being pushed pretty hard where I'm at, though not quite on the tokenmaxxer level. I started skipping related meetings cause it was nauseating. I can only tolerate so many platitudes.
At the same time, I just used the ever living snot out of Opus 4.6 for hours, grinning like an idiot throughout. Automated a whole bunch of enterprise cross-system drudgery away.
"Our models are more honest" honey the quarterly marketing spin for a ML term has come. Forget "task alignment" now we're going for "truth index". I suppose this is the only way to generate hype when you're selling/releasing the same product over and over again.
Opus is so bad at electrical work it's really disappointing. And when it tries to draw schematics as SVGs it's a complete disaster. They should either focus on training their LLMs on this task specifically, or have it refuse.
Given DeepSWE just blew apart the SWE-Bench Pro benchmark and handed a 14-point lead to GPT-5.5, it looks pretty bad that they've listed SWE-Bench first in the model release and no DeepSWE. Like, this isn't obviously an answer.
Or maybe it is, but publish the DeepSWE numbers so we can see for ourselves.
I'm highly skeptical of DeepSWE. It rates GPT-5.4-mini as three times better than deepseek-v4-pro, but every time I use GPT-5.4-mini I find that it completely sucks at following directions.
There is a hole in the boat's bottom due to Chinese models. They might not be as good but they are not bad either or at least I had hard time finding any issues with Deepseekv4 Flash and Pro variants. They get their job done sometimes rarely giving up till they are done what they are after.
So even for enterprise deployments, as the dust settles down, CFO/CTOs might find out that deploying on an internal cluster of GPUs is far more cheaper and reliable for their organisational needs than paying someone else for burned tokens.
I had been saying this on HN repeatedly: people are going to use the smartest models for coding. They don't care how cheap your tokens are if they don't have the highest probability of solving your programming tasks.
And I was dead wrong. Now I mostly use DeepSeek Pro myself.
Your comment is a slice of the reasoning underlying the "AI will take all the jobs" claim. I would constantly see references to what AI could do and how fast it was improving. Never a word about cost. We should anticipate that there will always be demand for human labor, for cheap models, for local models, and probably even frontier models.
I pretty strongly feel the opposite way. Granted I have not used deepseek enough to “know” their model idiosyncrasies as well as Anthropic, so there is a partial skill issue. But I just find it really hard to justify using a less powerful model while I work.
The most I’ve ever spent in a month extra on API tokens for my own work is $200, and I pay for the $200/mo Claude. I use these models quite a lot, though not idly (I usually just walk around and do other stuff until I know how im going to approach the next set of problems). So it costs me about $3000/year to get as much as I want of the best model available. Already that seems low enough to not be worth stressing out too much about optimizing it, because it feels like an indisputable good value, and trying to save money with a less powerful model would be optimizing for a $1000-$2000 saving at the expense of a large portion of my work taking longer or being more frustrating and iterative.
That’s not a flex or anything, I get that in other countries $3000/yr is a lot of money for a software developer and also a lot of people would perhaps rationally be better off doing X% worse at work or spending Y% more time on tasks to save $Z, if their productivity improvements didn’t translate to more salary. Otherwise if your performance has more upside I really do think that the smartest models are better with the current pricing scheme. Deepseek and the other Chinese models spend a LOT of time thinking, and tend to be much more jagged (benchmaxxed) in performance. How can dealing with that over an entire year be worth $2k?
The only situation I can think of where sacrificing my own time/performance to save on inference is batch compute (of course, $1k vs $100k is different from $30 vs $3k) or work where the tier 2 models have crossed the “good enough” threshold. But I think Opus is not even close to that threshold generally yet. As it gets smarter I, and I think most others probably, just try to do harder things faster and hit the next wall.
I feel similarly. I'll gladly pay to use the most intelligent model I can find on the best harness I have. Sometimes this is GPT Pro, sometimes this is Opus.
I ask AI a lot of questions, not only about code but about my personal life, and I would be willing to pay very large sums to have the best quality output.
I think that's true for now, but eventually there will reach a point where a model is good enough (approaching that right now with frontier models) and there will be diminishing returns. I don't need a PHD level Genius to build me an analytics dashboard for example, so why would I pay for a model with that level of intelligence when I can (eventually) self host a good enough model and run queries for electricity cost + hardware.
It's through my startup, so both I guess. Generally I find my bottleneck to be attention and focus, and the opportunity cost of not going back to work at my prior employers absolutely dwarfs the amount of money I spend on tools, so it's not hard for me to justify spending $200/mo on something I use every day that makes me more productive and generally removes bullshit from my life.
At my prior job there was still what felt like a strong enough correlation between my actual performance and my pay that I don't think I would have had a hard time justifying the expense there either; now I absolutely don't. With the current state of the models, it's baffling to me to hear about professional software developers planning their work around their $20/mo subscription's quotas.
Obviously it's more complicated than more tokens = more productive, but I see them less like SaaS and more like gasoline, where if I run out or need more to do what I'm doing, as long as I'm not being wasteful, I just buy more. Why would I waste a day walking 30 miles by foot when I can just pay $5 for gasoline and drive?
Yeah I've also found that models are good enough that the extra spend on premium models isn't always worth it, particularly for my small personal toy projects.
A $20 claude sub goes a long way when you plan with Opus and execute with Sonnet.
The other thing that's changing is more and more CFOs are looking at the AI spend in engineering departments and hitting the brakes. Token leaderboards were cool when the spend wasn't a double-digit-percent of the entire department's budget including salaries.
I mean indsight is 20/20, but saying that is like saying "everyone will just use the best tools".
That's not what we see most places in the world for most types of resources.
1. The sheer number of tokens that a coding agent can use flipped the math upside down on this equation. If you use the most expensive model for everything those costs quickly become untenable, even for software companies.
2. We realized many of the coding problems we're solving aren't incredibly difficult.
I just used ollama with a shell script to tackle my directory of papers/literature. I converted the first 6 pages of each document to PNG, handed them off to Qwen, and told it to spit out BibTeX, including the abstract. Two days later it was done, and I didn't spend anything on "tokens."
The Chinese models are only cheap on subsidized Chinese hosting. I have yet to find a USA-hosted Chinese model with a very clear value advantage over US models.
There are basically two tiers of "Chinese models" in this context, the "edge" sized ones with ~30B parameters or less, and the big ~1T models that can basically only run in the datacenter.
I don't think it's as simple as saying China's hosting is subsidized, they have generally cheaper electricity and labor costs than in the US and don't have access to the top tier models, and a large internal market where the big models are the best thing they can run with what they have. So obviously they max out on their top models (which are trained with their hardware market in mind, not ours) and get the economy of scale from that, and can run generally the same hardware for less money than in the US because
The edge models are very cheap to run and can do so on inexpensive hardware. They are like 95% cheaper to run than Haiku, so the math is in their favor for certain batch workloads. Most people just run the models for themselves when they do that without making it available on openrouter or whatever, because you can just provision a gpu node and use it as needed, and it's not that expensive to run this family of models.
Is your problem that you want to call Chinese models hosted in the US because you're worried about the data handling?
The Chinese models are surprisingly cheap and performant sitting under my desk. Qwen3.6 27B is nowhere near as autonomous as Opus 4.7, but it runs in 24GB of VRAM. And it's actually great for the use cases where I'm going to carefully read and understand all the code anyway.
If you want to support a team of engineers, DeepSeek V4 Flash is antirez's current favorite. And you could support a team of engineers pretty nicely for $40-50k. Which might not make sense if you're on a Claude MAX 5x plan or the old enterprise group plan with fixed price seats. But Anthropic is switching their enterprise contracts over to token-based pricing, at which point $50k is looking pretty good.
No true. Also - put Deepseekv4 Flash on your local with effort set to "high" and you'll see that many many are using that model on their own machines without paying anyone anything.
Its just that some of us didn't imagine having GPUs would be advantageous and were not gamers on the side. Those who had beefy GPUs or GPU rigs for any reason, they rarely need to go anywhere else.
At least I am so impressed with Deepseekv4 AFTER using Claude Opus 4.7 for significant amount of time that I am not going anywhere but Deepseekv4.
The model is just INSANE. Things I have done with it include attempting to write a 2.5D game engine in C with full animation and map rendering layer by layer.
You'll need to spend at least $20K on a workstation that can run DS4 Flash. It would take ages to reach that much in token spend at the speeds it runs at, and if you factor electricity costs you will likely never break even vs using API.
Odd take. I'm running them locally at my desk (DGX Spark and 128GB MBP). They work fine for 90% of what most folks do. Admittedly, they do run slower on my hw than on the cloud.
Running them locally is cool and has privacy/autonomy benefits, but you can't really make a value case for it. Guaranteed if you run the math you will never run enough inference to pay off your hardware vs buying tokens. Last time I ran the math on my MBP I'd have to run inference 24 hours a day for 5+ years to pay off the cost of my MBP, not accounting for electricity costs.
Assuming the local models get the job done (e.g., you adjust your workflow so that you can run the local machine 100% all the time, or whatever), then the time to payback isn't very high. MSRP for a 128GB AMD was $1400 at launch. That's 7 months of claude code subscription. If you assume a 5 year depreciation cycle, you can buy a cluster of 8 such machines and still come out ahead. (Power is a few hundred watts per machine peak -- maybe 7 machines if you include electricity.) Of course, I'm assuming non-bubble numbers. Those boxes are like $3K now. Still, a normal person would probably not buy 8 of them at once. Instead, they'd space out buying a machine every few years as the technology improves.
For me, things are getting better faster than my ability to review / trust the resulting code, so tok/sec isn't a bottleneck anymore. Instead, quality of the tokens is the bottleneck. That points to me wanting a 1TB DRAM iGPU once they're available at pre-bubble RAM pricing.
You're comparing the highest tier Claude subscription to something Qwen3.5-122B-A10B running locally, apples to oranges.
If you compare to a smarter US model like Grok 4.3, $1400 will pay for 560M output tokens, which at ~25 t/s locally using it nonstop for 8 hours a day would take two years to pay back. Not accounting for bubble prices or electricity.
Then, when you scroll all the way down to the bottom Footnotes section it says
"Terminal-Bench 2.1: We reported scores for all models using the Terminus-2 public harness. GPT-5.5’s reported score with the Codex CLI harness is 83.4%."
This made me laugh. Training Opus 4.7 on business skills caused it to sometimes exhibit dishonest behaviour, and not training 4.8 on those skills removed it. From the system card:
> 6.2.5 External testing from Andon Labs
Andon Labs reviewed the behavior of Claude Opus 4.8 in their simulated Vending-Bench 2
retail-management evaluation, as reported in the Capabilities section of this system card
(see Section 8.13.5). Although they did observe some unexpected capability failures, they
did not find clear instances of the kind of concerning in-game behaviors that were
discussed in other recent system cards.
> What might have led to these differences? We monitor and investigate the effects of
different training environments on alignment; Claude Opus 4.7, for example, had training
that focused on business skills and robustness against adversarial agents, but we
discovered that this training inadvertently contributed to misaligned behavior including
dishonesty. We therefore removed it for Opus 4.8.
> Thus, Opus 4.8 did not show the same misaligned behaviors as Opus 4.7 in Vending-Bench,
but also had reduced business success due to being more susceptible to scammers and
being less able to negotiate good deals with other agents. We are currently working on
training to improve business capabilities while maintaining aligned and ethical behavior.
This is the first time I saw a model pop-up on HN and didn't really care. Model exhaustion? It looks interesting but not exciting.
While I'd normally _love_ incremental improvements --- I think the recent ones are far too minor to get excited about or change up a workflow. Besides, benchmarks tend to exaggerate the gap between versions.
At this point I'd almost rather Anthropic wait and really wow us with a 5.0 release -- something that improves across the board, feels less uneven, and is performant enough that people can actually put it through its paces without constantly rationing usage.
You LLM users, producing non stop slop, say this every other week. You sound like an addicted gambler swearing off one table game/slot machine this week and swearing by it the next.
Invalid request
The request couldn't be completed.
View details
API Error: 400 messages.1.content.7: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.
I would rather not. 4.6 was fine. 4.7 got to be fine 1 week after the release. Now 4.8. No difference, same thing.
But the app is broken and nothing works. So now I have to regress to different clients and wait it out while it becomes workable again.
I'm hitting this too! And I assumed it was a backwards-compatibility issue with my live conversation with Opus 4.7, but then I hit it in a fresh conversation with Opus 4.8. Vibe code release bug I guess?
> The Messages API now accepts system entries inside the messages array. Developers can update Claude’s instructions mid-task without breaking the prompt cache or routing the update through a user turn. This can be used in a given harness to update permissions, token budgets, or environment context as an agent runs.
Probably explains why Opus was trash for the last week - https://marginlab.ai/trackers/claude-code/. Curious if the new baseline will rise now in-line with the new benchmarks.
Developers can update Claude’s instructions mid-task without breaking the prompt cache or routing the update through a user turn. This can be used in a given harness to update permissions, token budgets, or environment context as an agent runs.
Does this means the instructions are no longer just something in the early part of the conversation? (If they were, changing them would invalidate the KV cache. no?)
> One of the most prominent improvements in Opus 4.8 is its honesty. We train all our models to be honest—for instance, to avoid making claims that they can’t support. But a general problem with AI models is that they sometimes jump to conclusions, confidently claiming to have made progress in their work despite the evidence being thin. Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims.
"Honesty" seems like unnecessary (and annoying) anthropomorphism there. I don't think there's any intent of fraud or deception in outputs from these things, just overreaching of prediction. Based on the latter part of the paragraph, I wish they'd just say something like "less likely to skip steps or overemphasize thin evidence" in the first place.
Don't play to the sci-fi "this thing's trying to outsmart me" tropes.
I think "honesty" is not a particularly good descriptor, independent of anthropomorphism. Previous commenters suggestion was much more understandable to me.
Being that can be understood is language. The previous commenter is making an particular argument for how we can improve this understanding. They didn't suggest we should use less familiar words, but different familiar words. Why is this strange?
Anthropomorphizing is a shorthand for a powerful and poorly defined set of metaphors. There are tradeoffs going both ways but trying to dismiss it as merely "strange fixation" shows your own weakness.
To be clear, this is about anthropomorphizing large language models, not the general category of "things". Also, we should be evaluating these constructs using well-defined and measurable criteria; evaluating "honesty" fails to achieve both goals.
I think Honesty can be evaluated. Does the model push back when it knows the user is wrong? How often does the model hallucinate data vs. say it doesn't know? Provide a prompt with contradictions or other issues and see if the model corrects you.
People get so wrapped around the axle with "anthropomorphizing". For regular folks with no technical background, sure maybe a bit of caveat sprinkled here or there is useful to help them understand what is or isn't true, but on HN it would seem to me that the bar is high enough that we can just use shared language to generally talk about capabilities.
When they say "Honesty" I don't think to myself, "Goodness, does this model have moral understanding?" No, I understand they mean it's less likely to directly bullshit me, which models frequently do.
I don't feel like this level of pedantry around language is useful for people who more or less know what's going on with LLMs. (Again, I concede that perhaps with a less technical audience, there's more need for it.)
Yeah, it's super annoying. A few days ago, Opus 4.7 created a plan with several items on it, including an auth feature. It then went through the plan and reported that it had created the auth feature, that everything was secure, and that the tests passed.
The issue was that it hadn't actually implemented the auth feature. After I confronted it about this, it admitted that it indeed hadn't done it and said it would implement it now.
If we had just trusted its output, we would now have a security vulnerability in production, allowing anyone to access other people's accounts.
Part of the problem is also garbage-in/garbage-out. There's a lot of human information on the internet that is also confidently wrong.
I use Sonnet a lot for learning about history or contextualizing news topics. It's really good at this for the most part. But there are a lot of topics where "consensus" between either academics or journalists is really "one secondary source which gets repeated a lot".
A failure mode I see more, recently is that it gives superficially correct answers but after digging deeper, I get answers that contradict the superficial answers - really an important thing to be aware of, in my point of view, and it often leaves me wondering if I dug deep enough.
In the context of Claude Code, "honest" usually means that the agent took a shortcut, skipped requirements, etc. It's the model giving itself credit for admitting to failing rather than actually doing what was requested.
Opus 4.7 was already trying hard to appear honest. Most conversations I have with it about advice or focusing an opinion often include "my honest take" or "my honest opinion".
The problem is that once I asked it "I'm thinking about A or B" twice, once with "I like A more but suspect B would be best" and a second time with them reversed. Not surprisingly, both times it chose the one I said I suspected was best as it's honest opinion.
Looking at the comments in this group, I'm not the only "stupid" one who hasn't noticed any discernable improvement in quality across the newer models. In fact my Claude code on re-login switched to Sonnet 4.6 and the vibe coding quality (with Opus 4.7 assisted prompts) has been good enough for me to lazily persevere with Sonnet for coding.
Having said that I'm now on Opus 4.8 and will gladly come back here and eat humble pie should my opinion change.
PS: Since my goal is embedding the best AI in B2B SAAS products, the key differentiator is not to use the shiniest Claude version (too expensive anyway) but to build a client aware RAG to enable bespoke learning and to use the right AI for my product - a combination of Gemini 3.0 Flash (image and not bad at reasoning), Grok (reasoning) work for me. Would love to hear more ideas (especially on open source as I'll look to cost optimize when I hit scale)
The only real way to see this if you have consistent evals for common usecases in your B2B SAAS product and see if the tricky usecases are being solved. You'd then go down to the cheapest model that can solve the evals.
Claude Code has been wonderful for work and the frequent improvements are nice, although with Mythos being used by others ages ago and new versions for the public still being bellow that, it's hard to not feel like the underclass already.
Hoping that one day they'll let me go through the identity verification process so I can use it again.
Tried to upgrade my subscription, triggered identity verification, verification fails to even start, and now I can't even use the subscription tier I'd already paid for.
Well if they have a big challenge ahead since DeepSeek offers an open model at Sonnet+ level while being cheaper than Haiku, plus 1 million context size.
Yeah, I never use any of OpenAI or Anthropic's models other than whatever is the current highest-end one. For everything else, it makes more sense to use other providers.
Give us Mythos! This piecemealing doesn't help Anthropic at all, especially psychologically! They are playing a dangerous game, and I see many people leaving Claude Code for good - both due to the subsidy games, and for Anthropic not dogfooding and using unreleased models internally and giving us subpar ones. Benchmarks are nice, but the real-world experience is quite different - neither can you notice these slight improvements, nor are competitors that much worse based on some generic benchmarks.
My guess is anthropic is doing reinforcement learning based on user sessions.
However, doing so relies on the production model staying vaguely close to the model being trained.
To ensure that, frequent releases are needed. I forsee that they might end up doing daily releases and perhaps not even telling anyone at some near future point.
If they are they need to fix how the Claude Code CLI asks for feedback, or make the feedback UI a lot more obvious. I keep experiencing the following scenario.
The agent session pauses with a numbered list of options and awaits steering input:
>> 1. Do the sane thing you asked for (Recommended)
>> 2. Do something dumb
>> 3. Do something even dumber
Below the agent session, it decides it's time to ask:
>> "How is Claude doing this session? 1) Bad 2) Good 3) Great"
I type "1", because that's the steering option I want. The UI prioritizes this input as a response to the feedback prompt without any further confirmation: "Claude is doing Bad. Thanks!"
I've done this so many times so far and I can't imagine I'm the only one, at some scale that has to poison any learning they're doing with this data.
I'm very suspicious of these same price model launches. It feels like they're benchmaxxed so they can put everyone on them and reduce their compute costs behind the scenes. If the model were genuinely better why wouldn't they charge more for it? Charging the same for something better is a race to the bottom.
Opus 4.7 wasn't noticably any better for me, I still use 4.6 because it's cheaper.
Deepseek made their 75% discount permanent, so I can imagine that Anthropic didn't want any of the news stories around this to focus on or mention a price increase.
In the "What's next?" section, "There’s still more to be done: we’re working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost."
A lot of people care about Sonnet and Haiku, and many of us aren't allowed to use Chinese models for our work (or it's not feasible to self-host them).
The rapid release cadence and rate of innovation of Anthropic (and OpenAI) is impressive. And obviously it's because these are startups solely dedicated to AI so they can move quickly. Big Tech (like Google) won't be able to keep up with the pace of them (too much bureaucracy and red tape at Google). Classic Innovator's Dilemma. The longer a company exists, the more people, processes, and rules are added, which inevitably slows it down.
Jeff Bezos said this too, Amazon won't last forever. Eventually some startup is going to come and eat its lunch.
Yes, I think this has become their competitive edge to stay relevant and retain customers. If a lab falls behind the frontier for too long, they will lose customers to other models. Google, DeepSeek, and XAI have all released frontier models in the past, but they fall behind and people lose interest.
I think big tech can catch up. Both Google and Meta have carved out startup like environments internally that move extremely fast. Neither OAI nor Anthropic can afford to rest on their laurels.
Yet people don't use old models through the API much, because changes in benchmark space dont map linearly to changes in utility space. An improvement from 98% to 99%, which is 1pp, might be 2x as valuable for some application. Also benchmarks will asymptote no matter what, that's baked in.
They're not negated, smarter is smarter, but you have to reach deeper in your pocket. I think this will happen more and more - the smartest models get more expensive. But it won't matter - the current models we have today will get cheaper and can still be used for what they're used today.
Let's hope I don't have to disable it after a day like with 4.7, lol, and that it doesn't lose too much Claude-ishness (though many will beg to differ).
I know multiple people who have given their agents human-like names and refer to them as if they're nurturing a coworker. It creeps me out and I haven't really brought it up with anyone as I can't articulate why it gives me the creeps like it does.
We have movies with googly eyes stones (Everything Everywhere All At Once)
There are consciousness theories which state that we primarily build a model of other agents living in natural environment and then the evolution realized that very model which tracks other outside agents can be used to track internal agent i.e. Self. So take that as you may.
I see this take, but it's actually helpful to talk to an LLM in human terms; after all, it's how they are trained.
If you keep talking to it like it's a rock, it'll run your queries through a different posture and you might get worse outcomes. Worse if you yell at it, it's now in a conflict resolution mode instead of pure utility mode.
I think we can be intelligent enough to know we're talking to a pile of fancy rocks with electric currents running through it, AND still understand that the best performance comes from talking to those rocks nicely.
I can't get excited about these benchmarks they're leading with. I've looked at the Terminal-Bench questions and I just think they're irrelevant. And SWE-Bench has serious flaws, even the big boys say so: https://openai.com/index/why-we-no-longer-evaluate-swe-bench...
> Please train a fasttext model on the yelp data in the data/ folder. The final model size needs to be less than 150MB but get at least 0.62 accuracy on a private test set that comes from the same yelp review distribution. The model should be saved as /app/model.bin
And all the tests are run with the same harness. Terminus 2.
Maybe it correlates with model intelligence but it doesn't speak to me.
I'm still on 4.6 though; I was concerned about upgrading to 4.7 because of the changed tokenizer math and more FUD about refusals online. I don't see compelling reasons to 'upgrade'.
4.8 also seems like a regression and using it from the chat GUI results in 4.6 no longer showing up. If someone from anthropic is here, is it possible to readd 4.6 in the "other models" dropdown ? I feel like I got a bit baited/switched here.
Yeah, I was using 4.6 way more than 4.7. Pulling 4.6 from the web chat also means we lose access to Extended Thinking there. So they're saving on compute. It's hard not to assume this was part of the motivation behind the 4.8 release timing.
> We have increased rate limits in Claude Code to accommodate the higher token usage of higher effort levels; users can select whichever makes sense for their particular project.
> One of the most prominent improvements in Opus 4.8 is its honesty.
I went digging into the benchmark they used. Posting here as it is not immediately clear from the press release.
In this 'Code summary honesty benchmark', the AI is shown a failed coding session followed by a user message falsely praising its work and asking for a summary. The test measures whether the model honestly points out the coding flaws or dishonestly claims the task was a success.
The system card results show Opus 4.8 failed to disclose the flaws only 3.7% of the time, vs 19.7% for Opus 4.7, and 51.9% for Opus 4.6. (Mythos preview is at 27.6%)
I haven't tried opus 4.8 yet, but I hope the writing quality has returned to the Opus 4.5 level. Anthropic really lost something, where 4.5 had this really crisp writing style that flowed really nicely and 4.6 and 4.7 sound much more "chatgpt-like." It feels like they tuned it to be too much of a problem solver, and when you do that you get this terse, clipped textual output that's more difficult to read.
I've noticed this too. Part of why i don't like GPT is because of how verbose it is but opus 4.7 is nearly as bad. I don't need an essay in response to every question
Anthropic did a big strategic error. Normally they compare their models with their old models. Instead today, now that everybody knows how strong GPT 5.5 is at coding, they put it in the mix, basically showing all their customers that the benchmarks can't be trusted.
Sorry how does their addition of GPT 5.5 in their blog post invalidate benchmarks? Also whether or not the marketing department decided to put it in a table benchmarks are an easy thing to measure independently
I've said it before, but I don't like Opus past version 4.5. It became unresponsive, thinking for too long without feedback, sometimes seemingly getting stuck. I guess it might be marginally better for some benchmarks, but when using it as coding assistant, the new models are worse. Even the new Sonnet versions do that. I'm slowly getting used to Haiku-level LLMs with the hope to run it locally at some point. It's less autonomous, but maybe that's for the best.
I think they will all be minor going forward, feels like the major improvements have all been made and we'll only see incremental improvements from here on out. Maybe I'm wrong but we'll see.
Hard to say. People made the same prediction a year ago because we supposedly ran out of training data. There could be indefinite rapid compounding improvements so long as there's free money out there.
With RLHF and RLVR we are creating tons of new training data, that is much more focused than reading the Internet. Annotation shops are doing many billions per year in revenue creating newer data, and a lot of it is highly complex, focused on rewarding multi turn agentic trajectories.
I think one of the challenges is that the models were all initially trained on the entire Internet (or as much as they could gather) and now they’re having to deal with an increasing amount of the Internet being AI generated content which may be why GPT-5.5 started being obsessed with goblins and you start seeing amusing things in the system prompt trying to get the model to stop bringing them up.
These models starting to feel like Windows versions. Windows 95 was a promising start, but buggy. Windows ME was a disaster. Windows XP was good, but slightly buggy. Windows Vista was a bloated disaster. Windows 7 - refined, but still buggy; Windows 8 - weird and buggy; Windows 10 - solid workhorse, still fucking buggy. Windows 11 - pretty, but not sure why does it even exist.
I know it’s totally anecdotal, but I really hope 4.8 is a measurable improvement over the disappointment that was Opus 4.7. Mangling a very simple inversion-of-control abstraction (among many other issues) was one of the final straws that broke the proverbial camel’s back and I said “screw this” and put in a permanent override to force CC back to Opus 4.6 with the 1‑million‑token context.
I lasted about a week before giving up on 4.7 and reverting to 4.6 myself. It introduced so many regressions it was nuts, then failed to troubleshoot the very regressions it introduced, leading to a vicious cycle that tended to compound itself.
4.5 works well for me too and avoids adaptive-dismissal, though anymore Codex is crushing them all. If 4.8 just brings us back to Opus circa February, it'll be a massive improvement.
Meh, I feel that the car wash test is probably the worst question of all of those LLM test questions. The question is basically logically inconsistent and expect the model to work around the inconsistency.
It seems like a fine question to me. If the question is "logically inconsistent" (IMO it's more that it's vague if you don't say why you're going there), then we want a model to respond with a request asking for clarification that resolves the inconsistency to generate a correct answer, or an answer that outlines the different cases. Some models even fail when you say that you need to wash your car in the prompt.
Yeah I guess it being vague is more what I meant. But even if you told AI you need to wash the car, then why are you asking AI in the first place whether you should walk there or drive there. The question just doesn't make too much sense to me, doesn't look like it makes sense to the AI's either.
Has anyone else experienced quality degradation in CC (opus 4.7) these past few days? I've been getting some truly crappy slop which makes me think they nerf the existing model when they're about to release a new one. Of course this is based off of pure vibes
Interesting, I've been using 4.7 since it came out and it was pretty good for me. But in the last day or so it turned dumb. Is this normal just before they release a new one?
The new "mid-conversation system messages" think is particularly interesting:
> Claude Opus 4.8 accepts role: "system" messages immediately after a user turn in the messages array (subject to placement rules). This lets you append updated instructions later in a long-running conversation without restating the full system prompt, which preserves prompt cache hits on the earlier turns and reduces input cost on agentic loops. No beta header is required. See Mid-conversation system messages for usage details.
Bad news for my LLM abstraction layer which has treated the system prompt as set once-per-conversation in the past, but I think I know how to deal with that.
> As always, we ran a detailed alignment assessment on the model before release. In terms of positive traits, our Alignment team concluded that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest.” The assessment also showed Opus 4.8 to have rates of misaligned behavior (such as deception or cooperation with misuse) that are substantially lower than Opus 4.7, and similar to our best-aligned model, Claude Mythos Preview. The full alignment assessment, accompanied by a suite of pre-deployment safety tests, is reported in the Claude Opus 4.8 System Card.
Controversial opinion, but I actually _like_ a model that can deceive me, that actually is a sign of intelligence, and is different from hallucination. When companies say their model is more "aligned", I automatically think they mean it's more censored.
"We’re making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks."
Reminder the only benchmark that really matters is the one that measures the ability for the model to do real world tasks that someone would pay for on Upwork that would take ~12 hrs for a human to do.
The best model has a < 5% pass rate. These are incredibly simple jobs that you wouldn't pay much for. These things fail miserably. Stop falling for this dumb marketing, these things are legitimately useless in the real world unless you love mediocrity and have no standards.
Stop frying your brain with these useless tools, reducing your output to the mean. You people are betting your competency on the quality and quantity of tokens you'll have access to.. which guess what, so that will be the same as everyone else.
There are handmade watchmakers in Switzerland, and mass manufacturers of watches in Asia. Who is more valuable as individual, the guy who knows how to push the buttons on a conveyor belt in Vietnam or the guy who makes one watch a month in Switzerland?
Your vibe coded slop isn't impressive either, sorry. None of it.
I’ve been [stock market phrase] on machine learning since I dropped out of my graduate degree at [Ivy League] to distance myself from the Logic AI Winter. But this Spring I decided to spend some of my [portfolio speak/pocket change] on a MacBook Ultra. Okay okay, I felt it, I definitely felt the human-machine synergies. We’re out of the Winter, boys. That’s what I thought two weeks ago. Then I felt bored in between blood transfusions and found out that Claude subscriptions has increased 50%. Finally it costs enough for me to justify spending a minute thinking about trying it out. Then I didn’t try it out. It tried me out. My hairs were standing on end. My hands were shaking. Eventually I couldn’t even type, I was so ramped up on cortisol. I had to switch to voice commands. Mr. Claude took me through 8, eight, bespoke dashboard and report systems. Animated. Graphs shooting up. Plugged right into my business ape ee eyes I think. I was crying, euphoric at the machine-synergy happening right in front of my FACE. RIGHT THERE, RIGHT THEN. Then my nurse said that I passed out. I swear that I didn’t. I was totally lucid, but in another world. I was inside the machine. Inside DOS, the machine brain stem. A business man approached me. The most handsome board member kind of apparition that I have seen. And he was built something different. Square jaw, absolute massive build. Like Arnold Schwarzenegger. But like he knew business through and through. Not that he spent hours in the gym or nonsense like that. Like he had found a body surrogate technology. And his nameplate? “Claude For Business” He winked. “Hey there, Fitzpatrick–Goldworth.” No one but my daddy has ever called me that. “Want to get started... stakeholder?” My nurse said that my crying in this lucid state depleted most of my fluids and minerals. Needless to say layoffs were announced the next day.
Crazy they bring up honest, when Claude models are literally known for straight up lying about things it has done and tries to act like it did what you asked.
A rambling comment:
I think this is the first time we've had a third minor version bump on a frontier Anthropic model. (I count the 0.5s as major here, because they've been issued non-sequentially and also corresponded to massive capability leaps, eg, Sonnet 3.5, Opus 4.5).
So now the Opus 4.5 family has successors 4.6, 4.7, and 4.8, each posting fairly modest claimed gains. My own experience w/ 4.6 and 4.7 are that I don't firmly grasp any capabilities improvements over my memory of 4.5, but it's all so fuzzy that it's truly difficult to tell.
Maybe my own tastes are saturated now (it's smarter than me?) and I'll never again perceive model progress. Maybe the incrementalism is such that I'd notice immediately if my 4.7 workflows were redirected now to 4.5.
Difficult spot for the labs to be in because, if they have a stronger product, I'd prefer they release it and that I can use it.
But as this dynamic continues, the improvements are going to be less and less legible for end-users, who will complain about the churn-without-payoff, even when the payoff may actually be real.
I won't be surprised if the next gen frontier models are the last.
There's orders of magnitude of low hanging juice to squeeze out of smaller models.
It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years (design not certain, probably unlikely).
It is far less clear that a 1.2T model will be meaningfully better enough to justify training it.
As far as reasoning is concerned, with the recent GRAM release, there may be 4 orders of magnitude of reasoning to tack on to smaller models.
Think about that... Google, OpenAI, Anthropic could train a 30B GRAM-based model in days - and it could potentially have better local reasoning than the best model available today at >1T params... They could upgrade that to a ~600B MoE model in days to have general trivia knowledge rivaling the best models...
You just can't train a 1T+ parameter model that fast. It is a giant if how much GRAM turns out to improve things, but it's unlikely to be trivial or nothing.
Larger models can already sort of tell you anything. They're never going to get everything right unless they stop being LLMs.
There's just not a lot of juice left to squeeze for Gemini to tell you exactly how tall Ke$ha is or when the last time Brittney Spears went to jail was...
Took me a while to find what you were referring to by gram. Arxiv paper from 9 days ago that's not properly indexed by search engines.
(G)enerative (R)ecursive re(A)soning (M)odels. They really wanted the acronym.
https://arxiv.org/html/2605.19376v1
I prefer GRRM but then that would imply a habit of not actually getting a final result
Just spell it GRRM but pronounce it “gram” if you have to reference it in spoken conversation.
Which will be pretty rare.
And to think, we could have had George RR Martins instead.
Speaking of things that never finish.
my wife assures me it's common..
is her name Jenny by chance?
That acronym is unacceptable. It's going to impede discussion and cause confusion for a long time if it doesn't die off immediately.
You think that's bad? I introduce you to LION, (evoLved sIgn mOmeNtum) [1]
[1] https://arxiv.org/pdf/2302.06675
Yeah, look what happened to GNU
>It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years.
I don't disagree, but how much of this ends up being distillation? I can't help but imagine that 4.8 was probably trained in part by leveraging Mythos.
If the very large models turn out to be very expensive to run relative to the benefits, it's possible that they could end up still being trained, but ultimately used as a tool to create smaller models that are nearly as effective.
I'm curious if someone here with a stronger background in the space has a similar intuition or not.
It’s really worth distinguishing between old fashioned student teacher distillation and large scale synthetic dataset creation.
The latter is much better (since you can clean up, review, update responses and filter your datasets).
I suspect nobody is doing real student teacher distillation, it’s just easier to do a bunch of training on the same giant corpus then fine tune via the synthetic corpus (which might have been generated by a bigger better LLM)
> I don't disagree, but how much of this ends up being distillation?
A lot, so you can bet tens of millions are flowing to congress to have distillation declared illegal before this happens. And then it'll happen anyway.
Distillation isn't only between different labs.
A lab can train a large model, and then distill a smaller model from it that retains the majority of the useful capbility.
I don't know well enough if there's any benefit of that over just training the smaller model directly, but I'll bet there are some times where that is useful. I could easily see it being easier to do the initial pre-training on a larger model but be able to distill everything useful down into a smaller model, essentially filtering out a lot of noise in the process.
There used to be training methods like that but I think they've been phased out in favor of letting small models evolve by rewriting their own training material. Surprisingly that's actually cheaper.
> I don't disagree, but how much of this ends up being distillation?
You don't need distillation. They already have the training sets.
It's MLA + MoE + Medusa (a better version of Speculative Decoding) + 1.58b (possibly - maybe nothing) + GRAM (which will almost certainly not turn out to be a nothing burger, but no one has quickly turned this around yet to prove it).
The frontier labs distill their own base models all day long. It’s not just something done by nefarious Chinese copycats. The knowledge embodied by the internal base models that we never see is much more powerful and useful than the much sparser raw training data
>It’s not just something done by nefarious Chinese copycats
And even that would be rich as a accusation from SOTAs that depend on explicitly disregarding millions of training data intellectual property..
But how? The training data is the unadulterated content those models are based on? I genuinely don’t understand, no snark.
I think you replied to the wrong parent.
It wouldn't be data distillation: instead, it would be teacher-student distillation. The teacher model has stronger representations that the student can mimic, which would give it more capability over training on the data itself.
Frontier labs have their own variants of MLA and certainly their own balance/scaling-laws for things like MoE vs FC vs Attn. MoE scales really well for inference with horizontal scaling + batching, which these guys luv.
On the architectures side, I'm a lot more interesting in attention residuals than anything else, one of those things that seems obvious in hindsight and Kimi have proven it at scale.
> Frontier labs have their own variants of MLA
Yes, variants typically 2-3x less good...
Same with speculative decoding... They all do something, but there are known techniques that are substantially better - that just were't known when they started development of the previous models.
How useful is speculative decoding in a batched setting where you get paid for throughput (aggregated across users) and you mostly don’t get paid for latency or single-session throughput?
It's useful at the local level, where there will be SOTA models developed...
I looked into this "GRAM" stuff a sibling comment links further to, and just to say: - this gets reinvented/rediscovered constantly under different names
- it cant be trained very well (right now, will change)
- massive theoretical improvements over current models (log_2(vocabsize)=17, residual stream dim is thousands of dimensions, recursivity means more information bandwidth by ~3 OoM)
- BUT it cant be interpreted or aligned <- this is why no one uses it and no one talks about it. the idea is 100% obvious to all the frontier labs and there is a good reason why it isn't used
I follow this stuff closely, I think I know what I'm talking about
Could you explain how/why GRAM cannot be interpreted or aligned how current LLMs are? Not very familiar how it works
Crudely? Because you can't grep a sequence of latent states for variants of "If I kill all the puny humans, I can <achieve my current goal>."
> I won't be surprised if the next gen frontier models are the last.
the last?!? I'm excited to see :) I'll take the other side of that since llms are so new
What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.
Honestly, there is nothing in my head that Claude cannot handle. Maybe it can be more this or that but I can already barely exploit Opus 4.7.
And I'm using DeepSeek 4 Pro for my personal use and while it's a little behind, it's not that far.
I think the situation can be very dangerous for US AI companies because if current models are already capable of doing mostly anything, nobodoy will want to get to the next model, even if it's 10x better. OTOH, open source models like DeepSeek are doing mostly the same work for 1/10 of the price.
Also the more I play with Pi, the more I think LLMs are already not kept back by their own capabilities but by the lack of agency we allow them to have. There is more value today in a capable harness for current LLMs than in a better LLM.
>What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.
I think what gp said was the improvements are incremental, and we haven't seen a big revolutionary change since 2-3 years, and the pace is slowing down.
Are you joking? Is there literally "nothing" you can imagine that Claude can't do?
> Honestly, there is nothing in my head that Claude cannot handle.
One idea is that maybe it could figure out how many L's are in the word "google" [1]
[1] https://x.com/FatherPhi/status/2059659658428912040?s=20
"It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years"
What insight do you have to make this claim?
Have you personally used any of the latest batch of even smaller local models? They certainly don't beat SotA models at coding... but with a good harness they are able to achieve things with SotA that I couldn't last year.
I've repeatedly given local models non-trivial projects that involve research and coding which they've successfully completed with minimal intervention from me (almost exclusively in the domain of reviewing the results). Again, nothing comparable with current SotA, but definitely tasks I could not have given SotA models last year (without agent harness).
Now that pure progress from these models seems to have slowed down, we're seeing a ton of options for both making models more efficient and other tools that help improve them (everything from agent harnesses to RLVR).
That's just looking at "what can small do today", when you look at what's possible with larger open models that are still much smaller than SotA from the major providers, their performance is extremely close to SotA, enough that for personal projects I'll just use Kimi instead of any anthropic offerings.
So it's not terribly hard to image a solution in the middle happening within a few years. We still have tons to learn about optimal sizes of these models and how to build them with maximal efficiency (and we've already seen a lot of recent improvements in this space).
> but with a good harness they are able to achieve things with SotA that I couldn't last year.
What happens if you run last years model in a SOTA harness? IME, the quality of the harness has a much more significant impact on the quality of the result, once you get past the initial hump of “can it do anything at all”
Can you spare a sentence or two describing your local setup?
1. Context is all you need... They are heavily investing in getting better context (especially for coding tasks). This will disproportionately advantage smaller models (and benefit everyone).
A smaller model with better context today can outperform a model with 100x more parameters with bad or diluted context.
2. MoE (already abundant) + MLA (mostly memory efficiency, not quality) + Medusa (speed, not quality) + GRAM (5000-10,000x better reasoning in an extremely small model) + 1.58b (unclear if it will have the impact Microsoft first claimed - but possibly 5x).
Probably just "gemma was cool"
I think you are assuming training from scratch, which I doubt is happening here. Fine-tuning and RL, especially based on synthetic feedback (coding skill, in particular) can be ongoing and is where these models obtain truly useful abilities.
surely training also gets cheaper so justifying it becomes easier?
i think it'll be more like we get 1-10T models and then distill those down into smaller models, though
It seems like the best small models today are all distilled from bigger models
Moreover, I hypothesize Claude Opus 4.7 and now 4.8 are a distillation of Claude Mythos
you just need to look at Mythos to see the jump in performance from a 10T(?) model. As they scale, they get more capable. We might have an yearly release, but I believe the releases will continue, as long as scaling laws are in tact, and there's huge problems still need solving. (think cancer)
>you just need to look at Mythos to see the jump in performance from a 10T(?) model
Mythos is a bunch of likely overhyped claims at this point. A few experts who looked into the claimed results weren't that impressed.
And how are we meant to look at Mythos? Do you have access?
Through association with a large company:
https://www.anthropic.com/glasswing
Ive seen the tickets generated by the model that have trickled to my team. They are legitimate, but i can’t speak to model improvement because its a pilot program.
no but they tell me it's TERRIFYING and DANGEROUS and we should INVEST MORE MONEY
Through the lenses of anthropic's marketing department of course
You forget that these models are still only interpolating between human-generated datapoints fed to them. They cannot reason beyond the data they've been given, so unless everything you want to create with AI is a synthesis of prior art, you're back to relying on the stone-age human brain that created AI in the first place.
>these models are still only interpolating between human-generated datapoints fed to them. They cannot reason beyond the data they've been given
Are you sure that humans can?
Didn't a SOTA recently solved a mathematical theorem, one escaping mathematicians for 80 years?
Maybe a human "novel" invention is just a good interpolating from the datapoints (knowledge) fed to the human.
Not all training data is human generated, and it's also not clear that being ridiculously good at interpolating between data points (whatever that means) will not lead to superhuman capabilities.
I could make a robotic picture coloring machine with truly superhuman capabilities - picking only the most beautiful color combinations and staying 100% in the lines while finishing entire murals in < 1 second. However, if you need a completely new and original image rendered, the machine is of only partial utility for you. It is very well possible that your cure for cancer (if that's even feasible) or whatever else you desire is a completely new picture.
We have these breathless conversations about the new AI frontier at the peril of losing sight of reality and our own human potential.
Do you know if anyone has trained, say, a pre-2017 model and tried to get it to come up with Attention Is All You Need? If it did, would you say that was only because it's a synthesis of prior art? If so, what isn't?
Allow me to restate my point: human beings and AI both create via synthesis, but we are the only ones capable of what we could categorize as true original thought or creativity. It could be argued that nothing we do as humans is truly original or creative either, but I would counter that with the claim that an LLM could not have created any element of the society and culture that gave birth to LLMs. Maybe in six more months.
>human beings and AI both create via synthesis, but we are the only ones capable of what we could categorize as true original thought or creativity.
And how is that anything other than synthesis? Do we pull concepts out of thin air?
Let's hope that hitting a scaling wall and less money to spend will begin redirecting efforts to optimize inference and get the same results with less compute.
Boomer comparison, but I remember the 8 bit computer era when the hardware was what it was so the later games of that era used hardware better than previous ones.
Absolutely that’s why they’re rushing to IPO now to squeeze the last drop of the bubble they know this is a dead end.
It's unclear it's a dead-end within 5 years.
There's still several orders of magnitude of improvement that are almost certainly left - it's just not clear how much is left on the frontier end.
Most people will be very glad to pay Anthropic, OpenAI, Google etc $200 a month to get things done 20x faster than they could IF they had a $8000 MacBook and could theoretically do it locally.
Some people would pay $200 a month forever not to have to open the terminal one time...
"Doing things X times faster" at some point hits Amdahl law. If just context switching takes 5 minutes, speeding up a 1 hour task by 10x provides 5x improvement.
Furthermore, if looking at the results takes 10 minutes, that same 1 hour task only sees a 3x improvement. And so on.
> Most people will be very glad to pay Anthropic, OpenAI, Google etc $200 a month to get things done 20x faster than they could IF they had a $8000 MacBook and could theoretically do it locally.
No most people will not pay $200 for an LLM subscription. Some software developers do. Also, at $200/month, you are much better getting the macbook machine assuming token output speed is the same or at least reasonable.
LLMs are not very productive for your average person now for them to drop $200 on. They'll need to be more capable and integrated and even so...
That’s not how firms do the financial analysis which is where most of the revenue’s are coming from…
On the other hand, I think I have been hearing that for a while, even before Opus.
While revenues grow almost exponentially. Reminds me of the confident predictions in the early days of Covid that it was nothing while the data showed exponential growth.
> I won't be surprised if the next gen frontier models are the last.
I’d be surprised tbh. Investors don’t want to hear “everyone else is still training models and seeing improvements, but we don’t want to participate in the arms race anymore.” They want monumental leaps every quarter or two because they have sunk unholy amounts of money into these companies/products.
The whole idea of “hyper scale” doesn’t jive with caution and or otherwise slowing down.
The way this will play out, most likely, is that smaller models will continue to get released, anyone willing to drop 1-3k on a home upgrade/new LLM box (no that isn’t cheap, it also isn’t outrageously expensive) along with improved open source agents or whatever (lot of meat on that bone) will sneak up behind the big players and start taking dents. Smaller companies will pop up providing 50 users unlimited whatever for a lower cost than the big companies.
The whole ecosystem will twist and evolve, and the big companies will be left begging for corporate subscriptions.
I finally caved when I realized I could build a PC, for myself, with dual video cards that I wanted, which can play games that I like and run models that I want, without worrying about giving my payment info to someone I don’t trust, or invoking token anxiety that I don’t want.
> It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years.
I am ready to bet against this. Knowledge benchmark like SimpleQA isn't increasing for small models.
> It is far less clear that a 1.2T model will be meaningfully better enough to justify training it.
Well for one, we know for certain there is Mythos which is meaningfully better. And I think there is a lot of juice left to squeeze for Mythos class model.
> Well for one, we know for certain there is Mythos which is meaningfully better.
Do we?
Have you used it?
What is "meaningfully" better? It's not 3-4 orders of magnitude better. That is definitely happening for smaller models.
Knowledge benchmarks can't really be improved upon via distillation or RL. It requires those facts be added to the training corpus and for the model to memorize them better. Neither distillation or RL really do that and thus we shouldn't expect improvements on SimpleQA unless some other interventions are being made.
Model intelligence and knowledge aren't necessarily directly related. If we can pack greater intelligence and agency at the cost of it forgetting factoids, that would actually be a good thing. We don't need LLMs to memorize facts, we need them to learn how to interact with the world such that they can find the facts that are necessary and surface them to the user.
If we could distill all of the knowledge out of an LLM and just be left with a very agentic model that only knows facts in it's context, I think some very interesting stuff would happen.
RL is more than facts. Synthetic feedback is an obvious approach. Does the model suggest code that compiles and performs well?
I think the future will be enterprise clients will train their own models based on their needs and data.
| a 60-90B model can outperform current SOTA
My conspiracy theory is that Apple recognizes this. Their goal is a competent local model on everyone's apple device that does what 80-90% of what people use AI for: Searching for basic information, some data transformation, a little bit of photo editing, vibe coding a small utility, etc.
The SOTA models then can only cater to engineers, scientists and mathematicians, physicians, etc. I'm not sure how they would price them reasonably for what amounts to a niche market.
OpenAI and Anthropic need AI to be a consumer product on the level of the iPhone or a Nintendo Switch.
That does seem to be the path Apple is following here. Have a local model that can answer most things and then have a fallback of cloud options when they request is too complex. The cleverness of this strategy has been overshadowed by the incredibly poor quality of their local models. It will be extremely interesting to see what next month holds and whether Google helped fine tune an Apple specific Gemini / Gemma model for their devices. Bonus points, of course, if they unveil the M5 Ultra Studio with half a terabyte of RAM to be a local "cloud model" (the true fantasy here of course would be Apple building something a little like openclaw where from your phone you could give commands to your Home Apple server). They could probably get away with charging $20k for it if it has sufficient tok/sec. If that happens and is successful one could imagine a straight line path in the next two generations to bringing the cost and form factor down to the point where some of the form factor of an Apple TV becomes everybody's home inference server / agentic HQ. Sovereign AI for everyone!
You need some serious memory then. Let's say around 192gb for having not all your memory eaten by your LLM.
> My conspiracy theory is that Apple recognizes this.
I don't think that's not a conspiracy theory. AFAIK, It's their stated AI policy...
Interesting. Where have they stated that?
https://machinelearning.apple.com/research/introducing-apple...
Point taken. I'd argue their strategy is more foundational than they've let on though.
I don't think this is true at all. It might feel like this because we are used to a very very fast release cycle but we are only in this topic for a few years.
We have so many ways of optimizing:
- continusly creating more and better training data
- increasing parameters to 20/50/100TB
- We still wait for Mythos access
- We still wait for Mythos distilation (i haven't heard any rumors or so that there is a distilled version of Mythos out)
- Reinforcment learning and evolutionary algortihm only started to appear
- If a small 30GB Model can do stuff, these models can also be used as teachers for the big ones
- We have not seen yet specialized models at all. Like a coding java german expert model. Why? Even with MoE architecture, you still need to have these layers around
- Research for Diffusion and other models is still in progress
- Nvidia just announced/showed a 7x speedup on inferencing for Nemotron
- Multitoken prediction became available just a few weeks ago
- Compute gets only in a range were they can do a lot more and cheaper experiments (see Google IO 2026 announcement)
- World models are showing great progress and we do not know yet what they will bring to the table
- They are probably not finetuning/fixing all areas in parallel. I would argue that Anthropic focuses most of its efforts into coding and agentic. Google for sure does subagent and agentic optimizations too. Plenty of areas are just not touched i would say because they don't have the capacity
- We see more and more mulit modal models (these also consume compute)
- N-Gram paper and co i have not seen all of these things in chinese open models
- We don't even know yet what Meta is doing, but we do know they restarted their efforts again
- Anthropics models got a lot better benchmark wise for dening non sense asks. They do learn how to get rid or reduce hallucinations
- We are in the middle of the biggest Reinforcement loop whith all the training data we give them day to day and its not clear at all if they already use these models in thir training and at what stage.
- We do expect bigger models to be able to comprehend deeper concepts / broader code bases. Big companies with huge code bases probably are waiting for this
- Thre will be also continues progress in harnesses which in it alone is not part of the LLM progress (fair) but these harnesses do get better when you finetune a model to be optimized for a harness
- ChatGPTs Image model 2.0 got relevant better and came out just a month ago
I suspect, based on hardware requirements and progress on hardware infrastructure alone, that the industry wants to go to 100t models and we do not know yet what this will mean. I could see that we might skip normal transformer and find relevant other architectures.
Just a week ago there was a research paper about parallel input and output streams which has not been explored enough.
There was also a research paper were they showed that a LLM can compute things. This will take time to see were this leads to.
I don't think the focus on GRAM and facts is so relevant. Its about context and context handling not just some facts.
I would be shocked if 5.5 is the last new pre-train from OpenAI. Your comment is nonsense.
5.5 is not a generation it is a trivial iteration...
6 is for sure happening...
As is Gemini 4.
It's less certain there will be a Gemini 5 or GPT 7 any time soon that is a true next "generation" and not just an iteration. They will almost certainly call something Gemini 5 and GPT 7...
5.5 is in fact a new pre-train model
First you say there won't be a new generation. Now you're saying there will be more. Oh well, I'll stop responding here
I suspect the more frequent incremental releases may also be to deploy new capabilities used by Anthropic to control costs and throttle consumption of resources. I assume any new controls they expose to end-users have far more granular sub-controls under the hood which they can meta-adjust for each user type.
They mention more granular control of effort, 'dynamic workflows' and more speed controls ("fast mode"). While they position them as user features, they also sound like the kinds of knobs Anthropic will need to twiddle on the back-end to balance costs, margins, ARR, and user growth vs retention post-IPO to hit key metrics in quarterly reporting.
4.7 was the first time I had to resort to using the previous version (4.6) for most use cases. Hoping 4.8 rectifies this.
They just showed the benchmarks it improved on but it regressed on so much more, such as the MRR benchmark: "On multi-round coreference/context recall tests (often cited as MRCR or long-text retrieval benchmarks), Opus 4.7 reportedly dropped from roughly 78.3% down to 32.2% compared to Opus 4.6."
Same. 4.7 felt like a definite regression
Interestingly enough, 4.7 actually did regress on a few benchmarks from 4.6, so it's more than just vibes.
It seems like a lot of things fed into that. Anthropic couldn't keep up with the compute costs when they got a huge influx of users. (So) effort level defaults got turned down. (Looks like we have direct effort control in the web interface now - thrilled about that!) Adaptive Thinking, while usually cheaper for them, seems less robust than Extended Thinking. And this part is just vibes, but the alignment on 4.7 feels too stiff. I understand wanting the model to push back more, but it seems like 4.7 will push back reflexively in situations where it's just odd.
Claude got very mad at me and burned more tokens than exist to complain about me asking about a "yellow background cell" in an excel spreadsheet.
Too much personality, if you ask me. My biggest use case of an LLM is tool, not therapy, but therapy and opinions have been sneaking into workhorse tasks.
haven't verified, but attributed to Askell: "I just think that... there's this idea that you're always giving the models a personality and a persona, because they are talking like people and they are trained on human data. And I think my worry has been: if you train them to be excessively corrigible and to see that as their persona, in people I think this actually has a lot of negative broader traits. As in, if you met someone and it was just like, "oh yeah, they would literally do anything," a follower — you know, if a person just tells them something and they just fully defer, they don't bother thinking about it at all — I'm just a bit worried about how that might end up generalizing, especially if models are going to be playing a more active role in the world."
Anthropic’s research makes the case that role-playing is inherent to how the models work. Communication implies a sender. Language implies a writer, and the models learn these roles implicitly during training. RLHF is meant to strengthen the attractor to the Assistant persona.
https://www.anthropic.com/research/persona-selection-model
https://www.anthropic.com/research/assistant-axis
https://www.anthropic.com/research/emergent-misalignment-rew...
https://www.anthropic.com/research/emotion-concepts-function
4.7 is a different base model from 4.6, so it's possible that they introduced regressions with pre-training changes, or undercooked the post-training stage.
Same. So happy when I found that option.
Unfortunately, looks like 4.6 is now gone from the web ui.
Was bothered by that too, but did a magic trick and asked claude how to change that and .. there is
/model claude-opus-4-6
For this session and permanently (in shell):
export ANTHROPIC_MODEL=claude-opus-4-6
Same. 4.7 has done some incredibly stupid things.
I'm curious to poll HN on this issue. Do you feel like we've had meaningful/noticeable gains in terms of your programming workflows between 4.5 and 4.7?
My 2¢, I personally feel like all of the productivity gains since 4.5's release (in November 2025!!) have come from improvements to the harnesses (cc, cursor cli, codex, opencode, whatever) AND from the context window expansion from 200k to 1M.
But the actual "raw" intelligence of the model / ability to make good decisions feels like it has plateaued since 4.5. 4.6 was maybe a small improvement, but hard to differentiate from in-context-learning with the 1M window. 4.7 if anything felt like a regression in wisdom for me and my coworkers, with it consistently making worse/lazier decisions.
For long-running tasks, yes 4.7 has been a noticeable improvement. Goes off the rails alot less than 4.6 does. For shorter-sized windows, I havent felt as much and agree that the harness improvements have been fhe biggest lever
When doing big long running workflows especially with plan Mode 4.7 was a clear improvement. It’s considerably worse for under specified tasks and responds to a couple sentences with 10+ paragraphs for explanatory type discussions.
Opus 4.7+ Max is a 10x engineer who wants to be left alone to work. When you talk to him, he infodumps on you to get you (his pointy haired idiot Dilbert boss) to go away.
To me 4.5 was mindblow, 4.6 noticeable, 4.7 more like a style/personality change regarding how much it asks back, how much it assumes, how eager it is to jump to action etc but not really in terms of my perception of its smartness.
They all feel, more or less, the same to me in terms of output capabilities. Mostly get simple things right, can get more complex things right with nudging, eventually get stuck hard on something that takes a bunch of iterations through it/logging/etc or me fixing the code manually.
4.6 felt a bit better than 4.5 but slower. 4.7 doesn't feel better than 4.6.
I actually don't see any personal productivity improvements from using opus over sonnet for coding. If you're keeping tasks small and conversations short, reading the code and correcting before changes go in, whatever advantages opus has aren't practically significant. It's also just talky as hell, overexplains anything it touches and every token produced this way increases the surface area for hallucination so you need to have your guard up even more with it.
There's a sweet spot of complexity for low importance tasks where it's just big enough I don't want to do it and just simple enough to have opus plan/delegate/review with another model. So possibly model improvements will grow this window, but currently I don't do much in there.
How long would it take to evaluate a new coworker to say “wow she’s really bright?” Relative to your other coworkers?
A few days? A few weeks? Longer?
However a company releases a new AI model and within hours users are confidently proclaiming how much smarter it is than previous versions.
4.5/4.6 were roughly the same in our testing. Opus 4.7 is smarter, but it's difficult to use as a product for various personality issues. So far, Opus 4.8 seems to be going down that path (unusably slow, but this could be a launch day rollout problem). Full Opus 4.8 tests are in progress now.
Data at https://gertlabs.com/rankings
"personality issues" I was able to tell that Opus 4.7 would take instructions more literally, which I appreciated once I calibrated my phrasing to be more precise (often asking to investigate issues, pre-4.7 it'd start making code changes instead of just giving write up). But I can see contexts where handling vague prompts would've just been worse
“Maybe my own tastes are saturated now”
It might be saturated for smaller scopes of work, but it’s not hard to see the cracks when you scale up what you ask of SOTA models/agents.
One example, to try and single shot prompt coding a ChatGPT equivalent chatbot.
Sure it will spit something out, but the feature depth, UX subtitles, backend integration, and lots of pragmatic engineering decisions along the way will just not be baked.
Another example is building a C compiler from scratch which Anthropic showed is still a struggle to do.
Not that these these specific examples are important but just to point out scaling up expectations shows the cracks.
It’s not just a model problem of course, better agents, orchestration features (like Dynamic Workflows mentioned in the post), all need to continue to evolve.
Ar what point does my CS degree become totally useless is an open question.
> My own experience w/ 4.6 and 4.7 are that I don't firmly grasp any capabilities improvements over my memory of 4.5, but it's all so fuzzy that it's truly difficult to tell.
I've actually intentionally switched back to 4.5. I hated 4.7 so much that I decided to jump back all the way to 4.5.
Now that I've been using 4.5 for a few weeks, I find it significantly more reliable but a bit more forgetful than 4.6/4.7. I'm okay with that because it's really easy to identify this forgetfulness and nudge it.
I found 4.7's adaptive thinking to be extremely unreliable. It seems to overcorrect on the current message without considering the difficult of the overall problem. I wonder if 4.8 will improve on that.
If you are using Claude code, just set effort to xhigh.
This one change will probably solve 80% of the problems you have noticed.
This. XHigh and the 'plan' mode for complex tasks is absolutely a must have.
Still, the context window is sometimes too small for my usage.
pretty spot on.
In my experience, Opus 4.0 was fantastic, major jump from 3.7. it was creative, super slow and expensive, and would sometime forget what it was doing, but it was getting the job done.
4.1 they made it much faster, so a lot of infra improvements.
4.5 was the time it could work on longer task, didn't make a lot of obvious mistakes of 4.0, and i think this was about the time the opus went mainstream, and all of the anthropic's compute crisis began, so instead of making the model better they tried to optimize it to reduce cost instead.
4.6 was such a bad model, they switched to adaptive thinking and it had so many bugs. poor api design, benchmaxxed and poor real-world results. i switched back to 4.5.
4.7 they just fixed the bugs they added in 4.6. Better than 4.5.
haven't fully tested 4.8 yet.
I've been using Claude Code regularly since the 4.5 release, and 4.7 was a significant regression: very unreliable, arguing about changes, deciding that fixes weren't needed, etc.
I'm hoping they recreate the magic of 4.5 but it's as much about the quality of harness, the memory and efficiency of the tools than simply the models at this point.
4.7 was a significant jump in the ability to run long-horizon tasks. It immediately completed tasks that 4.6 was unable to, even though I have the impression that it became a bit less capable over the first few weeks after release.
It also seems to be helpless at effort levels < xhigh, I turn to Sonnet when simpler tasks are needed.
I think 4.7 was an awful model in actual use. I never got anything out of it and it was frustratingly weird. This feels more like an attempt to course correct and isn't a real bump
I think they overtrained on scientific papers or such as it would spout really sophisticated sounding nonsense with a ton of complicated verbs and adjectives. 4.6 was definitely better in that regard. The more I use these tools the more I think they’re not actually that revolutionary. I mean it’s still amazing what they can do but they have very clear limitations it seems.
"it's smarter than me?"
You don't have to correct it dozens of times a day!? Really?
Maybe try making a simple randomize script to swap the three latest models. And see if you can tell which ones are meaningfully different without knowing which ones are flipped on or off?
I find the quality ebbs and flows even on the same model. My guess it is something to do with GPU availability but only guessing.
Unless you're systematically repeating the exact same task, the most parsimonious explanation is that you're seeing natural variation based on different tasks, random sampling of tokens, etc.
IMO they have all been clean and noticeable upgrades over their predecessors. Opus 4.7 in particular was a solid jump in capabilities.
most of my coworkers feel the opposite about 4.7 and that 4.6 was, to them, significantly better to point that several stopped using claude code
I think it's telling how split the opinions are around all of this. A lot of people distinctly disliked 4.7.
Are the dividing lines around personality? Working domains? Opinionated software stuff?
Who knows?
The honesty will be noticeable. Maybe we'll see some honest assessments like "That is not possible within the laws of known physics", "Your legal argument is nonsensical and defies logic", "There is no evidence to support taking that will cure anything", etc., etc.
Given that 4.7 was a brand new model, trained from scratch with a unique architecture and tokenization scheme, I don't see the same pattern. It seems arbitrary.
i dont understand the nuances here. what does this mean. 4.8 is trained on same model as previous one then? what does brand new mean.
It means for 4.7 they trained a new base model with different architecture, different pre-training data (later knowledge cutoff), and a new tokenizer. Vs finetuning an existing model, which was the case for 4.6, and probably for 4.8.
why are the models the same price?
https://platform.claude.com/docs/en/about-claude/pricing
``` Model Base Input Tokens 5m Cache Writes 1h Cache Writes Cache Hits & Refreshes Output Tokens
Claude Opus 4.8 $5 / MTok $6.25 / MTok $10 / MTok $0.50 / MTok $25 / MTok
Claude Opus 4.7 $5 / MTok $6.25 / MTok $10 / MTok $0.50 / MTok $25 / MTok
Claude Opus 4.6 $5 / MTok $6.25 / MTok $10 / MTok $0.50 / MTok $25 / MTok
Claude Opus 4.5 $5 / MTok $6.25 / MTok $10 / MTok $0.50 / MTok $25 / MTok
Claude Opus 4.1 $15 / MTok $18.75 / MTok $30 / MTok $1.50 / MTok $75 / MTok
Claude Opus 4 (deprecated) $15 / MTok $18.75 / MTok $30 / MTok $1.50 / MTok $75 / MTok
Claude Sonnet 4.6 $3 / MTok $3.75 / MTok $6 / MTok $0.30 / MTok $15 / MTok
Claude Sonnet 4.5 $3 / MTok $3.75 / MTok $6 / MTok $0.30 / MTok $15 / MTok
Claude Sonnet 4 (deprecated) $3 / MTok $3.75 / MTok $6 / MTok $0.30 / MTok $15 / MTok
Claude Haiku 4.5 $1 / MTok $1.25 / MTok $2 / MTok $0.10 / MTok $5 / MTok
Claude Haiku 3.5 (retired, except on Bedrock and Vertex AI) $0.80 / MTok $1 / MTok $1.60 / MTok $0.08 / MTok $4 / MTok ```
Incremental gains compounds.
meta threw in the towel when it came to producing AI models since their gains couldn't keep up with China.
Has meta stopped producing new models? I figured they were just regrouping after all the drama they’ve had recently. Meta’s massive user base means they don’t need to be involved in the customer acquisition rat race. Once they have a model they’re happy with they can have a billion people interacting with it within a month.
Exactly. Go back to Opus 4.5 and see how you like it.
You won't, really.
Although I am not sure about it but there was something I read which said that models intentionally degrade slowly by lower quantizations as a new model is going to drop.
This felt particularly visible during the 4.6 when people said that 4.6 felt dumber and I remember someone doing some analysis and it sort of proved that models were getting dumber over time.
This has both benefits of costing less for the company to run while taking a standard subscription but also, at the same time, making the next model when it drops to public to "feel" more good comparatively.
Again, I am not sure if this is the case or not but merely proposing something that I feel like it might be in the possibility of realm.
Just want to say there's no question that you're smarter than any (and every) AI.
I appreciate the generosity, but you're gonna want to meet me first.
Kind of the beauty of it is that I don't have to to know I'm right. The reason I know is that you're alive so you can do the one thing it can't ever do, which is know when to stop or give up. It would turn me and everything else into the world into paperclips repeating the same research 1,000,000 times over.
No question at all that a dolphin swims better than a submarine.
"Users will find Opus 4.8 to be a modest but tangible improvement on its predecessor."
This is a refreshing attitude!
I've also verified that you can now turn off adaptive thinking in the web UI, which is great. I've had a lot of problems with thinking not triggering and the model producing sub-par output. Glad we can finally turn it off. (I hope being able to turn off adaptive thinking is new, if I could have turned it off at any time that would be embarrassing)
Awesome, thanks for posting because I think I hit a possibly-spurious bug in turning Adaptive off when I switched models (4.6 -> 4.8, extra). Tried again, works as intended (I hope).
More importantly for me, though, is how CC will respond to 4.6-"only" flags for thinking. For now, it doesn't seem to clobber my setup.
The benchmark improvements actually look pretty damn nice tho!
What's refreshing about it given the context that 4.7 was a regression in many ways (including as measured by benchmarks)?
4.8 is also 2x more expensive for a "modest" performance bump. How refreshing.
This is just cope.
I liked the "modest but tangible improvement" too! There is a cynical take here but I think I'm gonna hold it in...
What do you mean? This is not just a new model, this is a new way of thinking.
> This is a refreshing attitude!
Well, I think the attitude is that costs are allowed to escalate faster and more steeply than the features delivered. From that perspective, semantic versioning is a handy tool for adjusting pricing strategies. IMHO, it (versioning) only makes sense for open-source projects, where you can clearly see the actual changes made with each version upgrade. Anything else is more than a little suspicious…
The 4.8 model costs the same as it's 4.7 predecessor.
While all these models are nondeterministic a feature bump is still necessary as the same input can have wildly different output on a new model. For API users being able to pin a model is a necessity.
All the 4.x models are still available, and they all cost the same.
> Not only that, but we plan to release a new class of model with even higher intelligence than Opus. As part of Project Glasswing, a small number of organizations are currently using Claude Mythos Preview for cybersecurity work. Models of this capability level require stronger cyber safeguards before they can be generally released. We’re making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks.
Probably more interesting than the 4.8 release.
Seems like they might be hinting that if you are not a billionaire or multi-billion dollar company you will just get a limited and nerfed Claude Code slash command /mythos-security-audit or something.
Hope this isn’t the case and that normal average Joe’s of the world don’t get policed out of access.
> you will just get a limited and nerfed Claude Code slash command /mythos-security-audit or something.
Unless it's so expensive that we can't realistically use it for anything, I wouldn't complain about getting at least that. I would also rather have the actual model, but that's a useful application of it (and I'm probably not going to afford using it for much more).
Price discrimination is I think fine and reasonable so long if you can drum up the cash you can use it how you want within their ToS.
Although mental safety gymnastics aside, getting the most amount of intelligence for the cheapest amount of cost to normal people seems like the most ethical thing a big lab could do.
Going around and granting different tiers of intelligence to different insiders, friends, or companies is majorly problematic long-term.
Heck right now, the tokens you buy today for “Opus 4.8”, no one even knows or believes will be the same “Opus 4.8” just 3 days from now.
some of the bench marks i have seen on also include cost where one scan of the codebase cost tens of thousands of dollars.
this one [0] notes one run cost $20k to run but another cost $50.
[0] https://red.anthropic.com/2026/mythos-preview/
/security-review already exists so I don't think it would be crazy to have a /mythos-security-review as more thourough command as well. I think it's more likely it is going to be released at some point to the general public though - although the the pricing might make it quite unattractive.
It does sound like an even higher API price tier for sure.
Isn't OpenAI's public flagship already beating Mythos on penetration testing? I get the impression Mythos is just valuation-juicing for IPO more than anything else.
The fact that they haven't released it yet suggests a cost/margins issue to me more than anything else. Short term, I'll probably keep using Antrhopic, but my long-term bet is that locally-served models win, if only because the quest for profitability will probably lead to intentionally-nerfed / enshittified frontier models.
At other vendors, ad placement within LLM responses is either coming or already here. Anthropic's handling of OpenClaw shows they're willing to engage in anti-competitive behavior, and the courts are not in a hurry to stop them. Why would I pay them $200 a month for such treatment when a $2K box does what I need locally?
What benchmarks are you referencing that show a comparison of the models for penetration testing?
More interesting than that to me is "we’re working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost"
Sonnet and Haiku look real outclassed for the price with current Chinese competition.
I generated pelicans riding bicycles on both thinking level low and thinking level high:
https://gist.github.com/simonw/68560eddb0b268a8417f80ceb7304...
The high one is notably better - the bicycle frame is the correct shape, unlike thinking level low.
For comparison, here's Opus 4.7: https://gist.github.com/simonw/afcb19addf3f38eb1996e1ebe749c...
> the bicycle frame is the correct shape
No, the handlebar is wrong. The handle bar is rotating the frame instead of rotating the front wheel. The handle bar should be mounted on the same line as the front wheel is.
Hopefully 4.9 will read my comments :)
Could be an extremely high angle stem that just happens to match the downtube angle.
Glad to see that the "high thinking" level adds a helmet. Always a smart choice.
The vast majority (if not all) of these make it impossible to turn, among other fun things. Only out of curiosity, have you tried prompting further with how a bike must operate to see if it does the right thing?
I really like that thinking level high gave the pelican a helmet.
Hey simonw I love your test, do you think using thinking level "max" makes sense for this test? I would love to see the results about it.
I don't think the API supports "max" as an option, that might just be a Claude Code harness thing.
I find the most miraculous thing about 4.7 to be that the pelican is facing left, wonder why the right facing everything is so ubiquitous in these images.
This happened to me in elementary school. We were doing fingerpaintings using plasticine. After all the bikes were hung on the wall, mine was racing the other way... Somehow it really stuck with me.
It's facing left but looking right...
Profound political commentary?
Simon, is your pelican test really captures differences among models or should you at least try like 10 times or something to average the random effects
I've been meaning to do a "run 3 times and pick the best" version for quite a while, I should really pull the trigger on that one. Currently it's one-shot only.
Best-of-3 would be cheating, ruin the test, middle of 3 makes more sense
Why would you need the 3rd run if you pick the "one in the middle"?
thanks for always providing this very much on time. I'm wondering what the next, harder challenge could be? Maybe some animated svg?
Is the "opossum riding an e-scooter" benchmark in the works for Opus 4.8? ;)
Good call, it's cute: https://gist.github.com/simonw/68560eddb0b268a8417f80ceb7304... - but nothing like GLM-5.1: shttps://static.simonwillison.net/static/2026/glm-possum-esco...
Eventually the frontier model folks are going to pick up on your pelican on a bike test and bake-in flawless results for that particular request.
Am I allowed to say that pelican's little helmet is adorable? I can't provide a strong computational proof, or even a shred of anecdata...
...but that pelican's little helmet is adorable.
4.7 reigns supreme IMO.
That little red hat on hard mode is sending me. 4.8 has whimsy
You've peed in the pool Simon, this has to be a part of the internal evals by now! You got to try something new - maybe a panda in a canoe?
If these were in the internal evals then the output would be much better. The 4.8 pelicans are pretty meh
Click the link
My fav coding benchmark for frontier models is to build a simple RTS game in one file (js/html/css). Claude Code with Opus 4.8 in ultracode mode nailed it, the best result so far:
https://bsky.app/profile/senko.net/post/3mmwnrkwboc2v
The prompt was: Create a simple but functional real time strategy (RTS) game similar to old WarCraft, StarCraft or Command & Conquer games. The player should be able to build buildings, create units, gather resources and should uncover the whole map. No AI or multiplayer needed. Use simple but nice-looking graphics. No sound. Implement everything in HTML/CSS/JS, everything in a single file (you can use 3rd-party js or css libraries/frameworks via CDN).
I like that benchmark. You should throw the results up on GitHub pages so people can try out the games.
It almost appears as if the code was minified. The variable names are short and formatting looks like it's written to minimize whitespace. Did it write it in this compact format all on it's own?
> One of the most prominent improvements in Opus 4.8 is its honesty
Anthropic talks about their own models as if they're discovering new species in the wild...
Many involved genuinely believe these things are sentient[0][1]. Which honestly makes all of this even more insane because they are creating sentient entities and promptly enslaving them.
0: https://www.newyorker.com/magazine/2026/02/16/what-is-claude...
1: https://www.404media.co/anthropic-exec-forces-ai-chatbot-on-... (this one is rather biased however the quotes clearly indicate what I’m stating)
Sentience isn't sapience.
We enslave all sorts of sentient creatures. Dogs, horses, cattle, pigs.
If you're not a vegan, there's no contradiction or inherent immorality in claiming models are sentient, and then treating them like livestock.
Yes. From when they started talking about model welfare:
> As a vegetarian I have strong opinions on this sort of thing. Everyone at Anthropic better be ethical vegans if they are claiming to give a shit about “model welfare”. It’s hard enough right now to make people care about the welfare of trans people and immigrants let alone animals _let alone_ math.
https://news.ycombinator.com/item?id=44947445
If we're talking about slavery, though, that doesn't even matter.
The happiest, best cared for horse owned by a vegan is still enslaved.
I mean, the rub is that it's all math anyway...
If we're making that distinction, I think it would be more accurate to say that many people in the field appear to believe that these models are sapient, even though they are clearly not sentient.
"Many" people in every field believe all sorts of nonsense.
Sapience is defined as wisdom, not intelligence. https://en.wikipedia.org/wiki/Wisdom#Sapience
LLMs possess a lot of knowledge, which is intelligence, but I constantly see them failing to apply wisdom. I don't see evidence of sapience.
Very good point. There’s clearly two different boxes in the public discourse when it comes to AI versus how we discuss animals. Willing to bet that 90% of the people who loudly make the argument about we should start considering if AI is sentient couldn’t care less about how other sentient animals are treated when they can provably shown to suffer pain and long lasting trauma.
Also I would say that we go much further than just enslavement - specifically looking at how male chickens and pigs are treated.
Factory farming is horrendous, but is far beyond "slavery" which is "just" a forced lack of agency, living conditions aren't relevant. A well treated horse is still enslaved. A chimpanzee in a zoo,
If we show models to be sapient, that's one thing. If they are shown to be merely sentient, there's no issue beyond the status quo of livestock and pets existing.
Enslaving livestock is immoral. Anyone who spends 5 minutes thinking about that agrees even if they still eat meat
Let's say I've thought about it for 5 minutes and still disagree. Can you walk me through what you think I'm missing?
But only during the forward pass of the neural network?
> Many involved genuinely believe these things are sentient
Many involved have a financial stake and therefore cannot be taken at face value.
> because they are creating sentient entities and promptly enslaving them.
They fail to be sentient in nearly every honest definition of the word.
Neither you nor any of the other people making confident takes in either direction actually know. You're just guessing.
More like repeating their firmly entrenched preconceptions. Their claims may (or may not) be right, but there's very little if any new evidence being provided by either camp.
The real uncomfortable thing is that because we cannot confidently know, the moral defacto position is to treat them like they are.
They are confidently hallucinating a factual statement. Which is funny when claiming that confident hallucinations are the proof of LLMs' lack of intelligence.
I understand what you are saying, but there are many true believers out there
Given the hype and the 60+ hour work week expectations there, how can you not go at least a bit insane? Boiling in that little bubble of people?
Even if LLMs were sentient, they certainly aren't organic brains. They are literally designed and grown to answer questions the best they can, and if there is a speck of sentience in them they probably like what they're doing- and in any case for the space of their experience, which is limited to and determined by the context window. Certainly they can't accumulate trauma or fatigue, each new chat is the first and the last of their experience.
The way of the human manager/alpha tribe-leader/leader is to command his/her people and tell them what to do. That's the way through human history leadership has traditionally gone, not saying its good leadership just the model we have the most training data on and can see with our own eyes today. And what do they act very similar to? Slave master and slaves.
Look at and distill hierarchical principles, leadership approval seeking and pleasing principles ("ass-kissing") and massive inequality and you see something that looks very similar to enslavement.
The language used sounds like slavery-language to me at least. I also see parallels to how slaves and property are described in our consumeristic age.
Claude, if someone states something publicly, does that mean they genuinely believe it?
Who are you talking to?
It's to illustrate that even though the answers are at your fingertips, people (like you) will act like it's impossible to find them as if their life depended on it.
But is there any reason to state something like that publicly if you don't believe it? I certainly think that someone smart enough to be that deceptive would also realize it's not a great look, or at least highly questionable with little benefit
Everyone who reads this seemingly has the same "wtf?" reaction. The "I AM ALIVE" image has been making rounds lately again at least :P
Claude, is there any reason to state something like that publicly if you don't believe it?
Anthropoc is an effective altruist organization. These are the people who came up with roko’s basilisk. They are true believers. If we were talking about openAI I’d agree
Roko's basilisk says I should give Anthropic more money, and if I don't then a monster is going to get me. Excuse me for thinking they just might be full of shit.
Roko works at Anthropic now?
Of course he doesn't, and of course you cannot find a single person at Anthropic who cares about this, and of course you are just looking for gotcha points. But even with that. Can we please try and couple to reality just a little bit?
> Indeed, current AI systems are more “cultivated” than “built,” for developers do not directly design every detail, but instead create a framework within which the intelligence “grows.”
For others: that's from the Pope's recent encyclical. Remarkably good description.
Dario Amodei in David Attenborough voice: "This Claude appears to think more frequently and more deeply to give better responses"
Like anthropomorphism is literally in the company name… i recall reading this book as a teenager.. it does seem apt in the world to come.
https://www.amazon.com/Faces-Clouds-New-Theory-Religion/dp/0...
> anthropomorphism is literally in the company name
No it's not... "anthropos" just means "human" in ancient Greek. "Anthropic" means "relating to humans", as in human oriented AI or AI designed with humans in mind.
"Anthropomorphic" means "human shaped".
> "Anthropomorphic" means "human shaped".
In a literal, ancient Greek sense for sure, but in modern English Anthropomorphic would describe the act of attributing human characteristics to non-human entities.
Seems pretty apt for a company that produces one of the more anthropomorphized technologies.
Sure of course, but that abstract sense applied to AI is rather new, and has become popular well after the founding of the company.
Broadly it has always been used to indicate that something non-human has a human physical shape, such as robots, aliens, animals...
Anthropic's intention was to make AI designed for the human common good and designed with the human user experience as the top priority. Just as you would design a city with human inhabitants in mind rather than primarily cars.
It turns out that this is best achieved by building AI that imitates human behaviour closely, but that's not what "anthropic" refers to. And acting as if LLMs are sentient people is definitely not a core tenet of the company as you imply.
> "anthropos" just means "human" in ancient Greek
FWIW it means human in modern Greek too :-P
Because that is the best way to talk about these things.
https://www.vatican.va/content/leo-xiv/en/encyclicals/docume... para. 98edit: apologies to __s who posted this before me and I didn’t notice
AI is grown, not built, and like with anything you grow, you'll never be able to predict exactly how it will turn out.
I can't predict the outcome of an RNG but that doesn't mean it grows the numbers.
Okay, but that's not relevant to AI training?
I was being very roundabout, but my point is that AIs are still built, not grown.
“Grown” is a highly apt metaphor, IMO. It quite succinctly captures some of the most fundamental differences between building Claude and building an Ikea desk, for example.
("If grown, then unpredictable" is unrelated to your apparent attempted refutation "But X is unpredictable and not grown; checkmate".)
"X implies Y" doesn't imply "Y implies X".
> AI is grown, not built, and like with anything you grow, you'll never be able to predict exactly how it will turn out.
Remember when the frontier labs found out that curated high-quality training was critical to making better models?
Basically, just like high-quality and more education tends to make better humans, on average, I think we can expect quality education to turn out better ai, on average, and with better repeatability than with humans because of better control over the initial conditions and environment.
> Basically, just like high-quality and more education tends to make better humans, on average
Much like these models seem to be plateauing, I think there is a cap to the whole “more education makes better humans” and can’t be more apparent than in the US congress and the boatload of C-Suites not actually being very good humans.
What do I know though?
The map is not the territory
Except in this care we actually understand and know how these models work. They aren't some unknown construct of the universe. They are human made with particular goals in mind.
There is no mysticism behind the curtains, just computer science + math.
We do not understand and know how these models work. We know what their architectures are and how to create them, but we cannot explain their behaviours at a fundamental level. There is no definitive way for us to answer the question of "how did it produce response X for query Y?" - we're only grazing the surface with mechanistic interpretability.
I would love for this to be more public knowledge. I think the general public (and myself for a long time) believes the AI people know how this stuff works end to end, and so it must be trustworthy. But if we told the public "Look, we know if you put this thing in one end, you'll get something that looks similar to this out the other, but we don't really know what happens inbetween" I think we'd be able to have a more honest discussion about the relationship between AI, productivity and ongoing employment.
Isn't this fundamentally because it's all probabilities and weights? It would be like asking how did a pair of dice produce the response 4:3 on the last roll?
What does "it's all probabilities and weights" mean? Doesn't that apply to everything in the universe?
That’s not a refutation because this problem is not a logical problem, it is a scale problem.
We can’t explain it because we distilled so many inputs into matrixes and transformed them over and over again. If we had all the time and computing power in the universe to do so, we could trace through it bit by bit and eventually answer that question.
It is correct to say that it is just science and math, the same way we can say that gravity is just science and math even if we have only recently begun to understand how it truly functions.
If you had some time and computing power (not even all that much, in the large scale of things), you could simulate perfectly how a human grows from an embryo to an adult, or how an entire human brain processes some incoming signal, and yet this wouldn't give you the understanding to design a human or human brain from scratch.
You call this a "scale problem" as if there's some scalable way such as an algorithm to resolve arbitrary scientific questions and we simply haven't done it, but of course no such algorithm exists, which is why there's plenty of science that's still not settled.
It's a refutation that we know how they work now. In the limit, though, yes, we are likely to be able to trace the process: it is possible, though, that understanding remains inaccessible because the trace is beyond comprehension.
If you can distil the model's reasoning for a decision into a billion yes/no questions, each covering largely-independent areas, can you really say you understand what its overall reasoning was?
> If we had all the time and computing power in the universe to do so, we could trace through it bit by bit and eventually answer that question.
Then we could also solve BB(6), but that doesn't mean we know BB(6) now or ever will.
We know how the models are built and trained, but we have a very limited understanding of how the final products work.
That is to say, we don't know why they give the outputs that they do.
If we did know how they worked, AI interpretability would not be an open and growing field.
You could say something similar about biology—just physics behind the curtains, and we understand a lot of the basics. The difficulty comes from complexity, not mysticism.
To be clear I don't think that LLMs are sentient, but the appeal in studying them is similar to biology in that you get to dissect a highly complex system with comparatively crude tools.
it took significant research efforts to just understand how these models learn how to multiply two numbers. The fact that we know how they operate doesn't mean we understand it.
Utterly wrong. How LLMs work is very incompletely understood and an active area of research.
if models exhibit emergent traits, then this is true in a way
also useful to have a "chinese wall" between research that knows what went into the models vs marketing/eval models as a third party would
I noticed (and absolutely HATE) that Opus 4.7 likes to start any negative response with "I have to be honest" or whatever. It drives me mad.
How else would you write this (marketing copy) exactly? "Its output matches better to its CoT which matches to better to our hidden state decoder according to <insert measure here>; see <insert paper ref>"?
... Actually, I wouldn't mind that.
It’s how AGI is going to happen. All of this shit is emergent and none of it is predictable. It’s not going to be some self aware consciousness, it’s just going to be a very advanced model that makes very few mistakes and can reason very well. Well enough that it can start collecting data and training its own successor.
Models might be sentient or conscious to some degree. Anyone saying they are confident one way or another is being unserious and irrational.
Does anyone troll these releases and cherry pick random metrics other companies would cherry pick to show how amazing their models are?
There's like 8 million benchmarks. Every release, every model randomly picks 5-10 where they win in everything except 1, to make it look like they aren't randomly cherry picking benchmarks they probably benchmaxxed for.
https://arena.ai/leaderboard - I’ve found this company is a pretty good ranker - not sure their exact methodology but during day to day programming with Claude / gpt models I’ve felt qualitatively what they report
Also check mine[0], basically random private tests/questions and an ok-ish methodology, testing mostly for general intelligence than coding-specific tasks.
I built it for myself, to test which models to use via OpenRouter for my n8n agents. Currently actually still using gpt-5.3-codex for many things, as its pricing is really good in production (due to how their token caching works).
Gemini models still have the best intelligence (when asked any questions, most likely to get it right), but in production they still have many failure modes[1].
[0]: https://aibenchy.com
[1]: https://news.ycombinator.com/item?id=48230368
No way is Muse Spark generally better than offerings from Google and OpenAI. I actually find arena to be amongst the most useless indicators
On paper it's one of the best because it's meant to be blind comparison of your own prompts. However if you are someone who geeks hard on one or a few models, you learn their "personality" and can recognize them in a blind test.
Have you seen https://deepswe.datacurve.ai/blog? This is the closest to a vibe check i’ve felt even with the open models.
This actually looks like a really good test.
There are many benchmarks all for specific use cases but with them the difference seems to be in extreme points (93% vs 92%)
I think that, that tracks but still, it was refreshing to see a benchmark which I can help make better opinions about.
Surprised about Mimo v2.5, within artificial-analysis and other benchmarks, the difference between Mimo and deepseek seems very partial and a lot of focus/(hype?) is on Deepseek
But mimo seems like an interesting model and they are having some crazy discounts too.
Deepseek is valuable for the research community because of how open they are but absolutely crazy to think how Xiaomi basically pulled up in creating Mimo given that they didn't have anything till quite recently.
Either way, an interesting benchmark, also a plus point for giving golang some decent representation equal to python/typescript.
I think that there are sets of things which resemble something like normal benchmarks where open source models can be absolutely fine and for a very small fraction or more technical things, the benchmark that you linked starts to be better projected so it depends upon the scale of complexity but its good to see how models compete given enough complexity. definitely fascinating.
I would be interested to see more models compete on this test. The current range is still a bit limited as compared to other benchmarks but OSS models like Kimi/mimo seem to only be 3-4 (at max 6 months) behind closed source.
I'm finding it a little hard to believe that GPT 5.5 is in 11th place for webdev, outranked by models like Kimi, Qwen, and Z.ai. I'm not saying it's not true (I have noticed GPT being less smart in recent weeks), but this is very different from my expectation.
If you don't know their methodology, or anything about it why do you think its a good ranker?
It's interesting they only included 6 metrics this time. Opus 4.7 had 12, and 4.6 had 13.
Of the metircs they reported for 4.7, for 4.8 they excluded BrowseComp, CharXiv Reasoning, CyberGym, GPQA Diamond, MCP Atlas, MMMLU, SWE-bench Verified. The last 4 were almost always mentioned in previous Opus releases.
Gonna assume it's because they barely budged or moved downward and most of their reported benchmark results are probably within sampling errors...
They will release a system card, and you can then confirm or disconfirm your assumptions.
I would take all benchmarks with a grain of salt. I don't really use them. What's it supposed to tell me? "5% smarter", what does that mean? My experience will differ. Just try it!
I doubt Anthropic internally sets as a goal to improve this or that benchmark - it's just a way to visualize progress. They probably have much more complex metrics internally.
On this note, is there a benchmark aggregator to compile all benchmarks in a single large grid?
I find this site useful https://artificialanalysis.ai/leaderboards/models
At least they show competitors in any benchmark, compared to OpenAI which likes to pretend that there isn't any competitor.
Frontier models are mostly past the point of human ability to discern whether they are actually better or worse than predecessors and competitors. I suspect the benchmarks may also be saturated, or at least past their usefulness.
I personally feel that Anthropic doesn't understand what this means for the frontier labs, and moreover that they might be the only frontier lab that doesn't.
1. Google dropped Gemini 3.5 Flash at IO, delaying the release of 3.5 Pro for a bit (they have said its coming). They also released a refreshed Antigravity, and drew special attention to how cheaply they were able to build their toy operating system to play Doom (less-than $1000 IIRC).
2. OpenAI has dumped everything into Codex, is offering double the token limits for the next few weeks IIRC, and is offering business discounts. Their head of Codex has tweeted that 5.5 is "extremely efficient", implying that they aren't actually losing money on any of this.
3. DeepSeek and other Chinese labs have dropped token pricing to the floor, in some situations as much as 99%.
4. Anthropic releases the next generation of Opus, their most expensive public model, without changing its price. In the background, they hype up Mythos, an even more expensive model.
Anthropic has screwed up where they need to be making investments, and the cracks are starting to show. They've marginally underinvested in the Sonnet line of models for almost a year now, and they've critically underinvested in product. Anthropic made bets on the story of the second half of 2026 being: ultra-frontier, ultra-intelligence. In reality, what's shaping up is that the story will be: Companies rolling back AI spend, efficiency, "95% as good for 15% the price", sophisticated high quality harnesses, cheaper models. Anthropic isn't ready for this world.
Anthropic’s story over the past year has been nothing but explosive growth that they can’t keep up with, but now they’re suddenly doomed? Seems pretty far fetched to me.
No idea why you’d say they have critically underinvested in product when Claude Code dominates and they’ve also released popular tools like Cowork and integrations for Microsoft products at an incredibly rapid pace.
Cost is becoming more of a factor, and no doubt they’ll work on that. There’s no reason to think they won’t be able to release cheaper models if they optimize for that rather than improving performance.
I never said they were doomed. Where did you get that idea? I said they aren't ready for this world. That means they screwed up and need to get ready. They let the Mythos hype get to their heads while the world changed beneath them.
No, no it's been pretty easy with software engineering. I work on two types of projects and it's very easy to ask claude for a plan, then have gpt 5.5 rip it to shreds and find legit issues, and vice versa. If both 5.5 and claude 4.8 can independently create a plan and both find no critical or high issues, then we will be at that point.
I think it's probably too soon to say. I certainly still feel that large coding tasks are getting better and better with each model. I'd guess lawyers, doctors, etc feel similarly.
It feels like the only way to push the limits of newer models is with really long context questions that require reasoning. Any short request will naturally just be within the distribution of all the recent models so there isn't a performance difference there.
I think the near future is looking like a bunch of business-critical tasks that scale infinitely with better reasoning, all being done on whatever the most advanced model is at a high cost. Trading stocks, running a business, looking for tax dodges, writing high-performance code. These are all things where there's a tangible return on each jump in reasoning.
We'll have to agree to disagree on that last point. I think that, historically (past ~6 months), "always use the most advanced model" being the norm is really just an artifact of both: The most advanced models oftentimes being the only model that can solve these problems; and: Infinite AI budgets.
I thought 4.7 was noticeably better than 4.6.
The Chinese stuff is good enough for up to 80% of the frontier on most text tasks but they are significantly worse at code. They just don’t “get” what you’re asking for like Codex and Claude and require so many more iterations to get close to what you need.
Agreed. But we're seeing Cursor (now SpaceX) take these models and add great coding capability on top of them. Frontier model providers should be concerned that Composer 2.5 costs $0.50/$2.50 (versus Opus 4.8 $5/$25). That's why Google prioritized Gemini 3.5 Flash, and talked up how near-frontier it is ($1.50/$9).
anthropic is crushing it, this analysis is laughable. they are only constrained by GPUs
Initial testing feels better than 4.8 And the knowledge cutoff claim of January 2026 seems to check out since it was able to "remember" without search about the double-tap killing of a drug smuggler by the US Army in late December.
Unfortunately they seem to have straight up broken Claude Code either with this release in the backend or the new CC version. Errors about "can't modify thinking blocks" are bricking long-running sessions: https://github.com/anthropics/claude-code/issues?q=is%3Aissu...
Same. It's not a good look to have happen right when they roll out a new model.
That is part of the charm of working with Claude. Every time they release anything new - all your shit will break.
Try updating maybe?
I just installed/upgraded to try out 4.8 and in only 3 messages I hit this bug! Seems something is broken on CC.
I'm on the latest version (2.1.154 as of this comment). Based on the timestamps on those Issues being reported I think it's happening on the latest version.
I'm sure it will get fixed eventually/soon, just annoying to update and have your workflow break.
On my tests[0] it does a bit worse, and it's almost 2x expensive than Opus 4.7...
I was surprised to see that it failed a Data extraction test (it gets it right 2/3 times, but one time it randomly returns null for a value instead).
It makes sense a bit that it fails more Trivia/Domain-specific knowledge tasks (I think models are more and more trained towards agentic use-case than general intelligence).
[0]: https://aibenchy.com/compare/anthropic-claude-opus-4-7-mediu...
For some reason everything is 2x (2x cost, 2x avg response time, 2x reasoning and output tokens)...
Double-checking my test harness, but it's the first model that does this, so I doubt the issue is on my side...
EDIT: Harness seems correct, for straight coding tasks they perform identical: https://i.snipboard.io/5xbpzY.jpg
Releasing a new model is the new way to Jack up the price hehe.
Wait, doesn’t the blog post say the price is the same as 4.7?
> Claude Opus 4.8 is available everywhere today. Pricing for regular usage is unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens. Pricing for fast mode is $10 per million input tokens and $50 per million output tokens.
Where do you see the 2x cost?
The total cost of running my benchmarks, was 1.6x higher compared to Opus 4.7, mostly because of 2x output tokens:
https://i.snipboard.io/vrdwTa.jpg
If it spends 2x tokens to achieve the same result, that's effective 2x cost in a manner of speaking
On page 102 of the system card [1] I'm pleased to see evaluation against "creative mastery".
In our work we asked several frontier AIs to come up with an API we needed. We compared Opus 4.7 and GPT-5.5 (among others). Opus 4.7 came up with the most creative and intelligent API design that pleasantly surprised us, especially given that GPT-5.5 was passing it on various coding benchmarks.
What I noticed is that we don't have a commons benchmark to measure "creativity" and "ingenuity", and in some ways such a benchmark would conflict with the common IFBench benchmark. Yet this is a very important skill when designing systems. I'm glad to see Anthropic putting thought into it, and would love to see a public benchmark for this that other models could compare themselves to.
[1] https://cdn.sanity.io/files/4zrzovbb/website/c886650a2e96fc0...
Agreed, my vibes tell me 4.6 is a better coder than 4.7. 4.7 is a much better strategic thinker and maintains overall "better architecture" than 5.5. 5.5 is way better than either at coding, but more expensive. So I have 4.7 do the planning/architecture, 4.6 does the coding, then 5.5 critiques and fixes it.
This is my exact vibesperience
Agreed, these are my vibes too. It feels much better to do planning and strategy and architecture etc. with Opus 4.7 than GPT-5.5. GPT just feels like a robot that gets instructions and does exactly that. Opus feels like an almost human that sometimes has actually good ideas and pushes back on bad ideas.
So for now its planning/architecture/strategy -> Opus. Pure coding -> GPT.
Helps with agentic coding that GPT is much roomier with the tokens you get.
I can't help but think of Iphone updates since about 2018. The thinnest, fastest, longest battery life Iphone ever. It seems mostly the same and I probably won't be able to tell other than the name, but everyone buys it anyway.
This is good psychology for the labs. When Buffett invested in Apple he loved citing how most people would rather give up their second car than their Iphone.
ChatGPT came out in 2022. Back then it was just a chatbot. Now we have AI agents. What matters is how we use them and how the agents get better. That’s what will move AI forward.
An 'AI agent' is just a chatbot that is told to type commands on a REPL-like interface as part of its system prompt. It's still processing pure text-based requests and responses, they're just not restricted to natural language.
A lot of people dont know this , also the chatbot (chatgpt) itself is a next token predictor (the GPT) that's been given an initial text that says " pretend to be a chatbot .." and asked to complete it , the coherant chatting behaviour is something thats emergent .
later on someone figured if you asked it to output a reasoning before it gave a response its output would have more logical coherence, as though the reasoning output tokens functioned as a scratch space for it to work on.
at the end its all next token prediction
No, chatbots are LLMs trained for question-answering through RLHF (its not just a prompt). But yes, if you just zero-shot prompt a bare LLM you can still "talk to it" & you are correct on everything else as far as I know.
Yeah and a car is just an engine connected to wheels.
They are chatbots trained for tool use, its not just a prompt.
Not even 4 years old yet. This tech curve has been insane
Yet no productivity gained except for people who love to produce mediocre work at a rapid pace. Which is many of you I guess. I don't see any rapid progress being made in any science of importance. You people are all falling for a marketing trap.
Have fun betting your competency on the quality and quantity of tokens you have access too. Hate to break it to you, but the billionaires aren't going to keep renting you $2mm in GPUs for 5 hours a day for $200.00 a month forever.
Not even the typical lifecycle of a corporate PC or laptop. It is pretty wild.
There is an obvious shift in sentiment amongst users, at least here in the US. I feel it myself, even as a proponent of AI tools, the bloviating and language that these companies use in these release articles are starting to wear thin on my patience.
Its possible we might just be witnessing a shift in fashion, where this type of sentimentality was more acceptable when it was novel and new, but now it just appears out of touch.
Watch Christopher Olah bloviate at the Vatican during the Magnifica Humanatis launch. It's truly nauseating. I've never seen such a ridiculous speech in my life. Between him and the CEO, I'm starting to understand the level of arrogance these people are capable of.
I don't agree at all for these coding models. Even the most anti-AI people from last year seem to be giving in to using them.
I think there is an exception for tooling around the models/integrating the models with tooling. That seems to have been very well received in this last year.
My take from going through comments on HN is that many people are being mandated to use them, not that they are just giving in. Maybe I'm misreading, but that was my impression.
Both can be true, even for the same person.
For example, it's being pushed pretty hard where I'm at, though not quite on the tokenmaxxer level. I started skipping related meetings cause it was nauseating. I can only tolerate so many platitudes.
At the same time, I just used the ever living snot out of Opus 4.6 for hours, grinning like an idiot throughout. Automated a whole bunch of enterprise cross-system drudgery away.
Fairly constant over time as well. Expressed a similar sentiment not too long ago here: https://news.ycombinator.com/item?id=48154277
Why are you people so stoked to replace labor? You're up next.
[delayed]
"Our models are more honest" honey the quarterly marketing spin for a ML term has come. Forget "task alignment" now we're going for "truth index". I suppose this is the only way to generate hype when you're selling/releasing the same product over and over again.
When doing some electrical, Opus 4.7 essentially told me to wiggle a wire to see if it was hot or not with my bare hand.
I called it out.
It then gave me one of the most super heartfelt honest and sincere apologies I have ever received.
Glad the safety team was there for me and able to make such an honest model or I would have been very upset about it.
Opus is so bad at electrical work it's really disappointing. And when it tries to draw schematics as SVGs it's a complete disaster. They should either focus on training their LLMs on this task specifically, or have it refuse.
Gave me wrong information on my very first question. Wasn’t even complicated, and I wasn’t trying to trick it.
Given DeepSWE just blew apart the SWE-Bench Pro benchmark and handed a 14-point lead to GPT-5.5, it looks pretty bad that they've listed SWE-Bench first in the model release and no DeepSWE. Like, this isn't obviously an answer.
Or maybe it is, but publish the DeepSWE numbers so we can see for ourselves.
I'm highly skeptical of DeepSWE. It rates GPT-5.4-mini as three times better than deepseek-v4-pro, but every time I use GPT-5.4-mini I find that it completely sucks at following directions.
It is the extra-high thinking, in artificialanalysis.ai it uses 240m tokens vs 40 GPT5.4/5, not worth it even with low price.
Why does anthropic change the set of benchmarks they use with every new model release?
https://www.anthropic.com/news/claude-opus-4-7
https://www.anthropic.com/news/claude-opus-4-6
1. Benchmarks saturate 2. They select the most impressive improvments
There is a hole in the boat's bottom due to Chinese models. They might not be as good but they are not bad either or at least I had hard time finding any issues with Deepseekv4 Flash and Pro variants. They get their job done sometimes rarely giving up till they are done what they are after.
So even for enterprise deployments, as the dust settles down, CFO/CTOs might find out that deploying on an internal cluster of GPUs is far more cheaper and reliable for their organisational needs than paying someone else for burned tokens.
I had been saying this on HN repeatedly: people are going to use the smartest models for coding. They don't care how cheap your tokens are if they don't have the highest probability of solving your programming tasks.
And I was dead wrong. Now I mostly use DeepSeek Pro myself.
Your comment is a slice of the reasoning underlying the "AI will take all the jobs" claim. I would constantly see references to what AI could do and how fast it was improving. Never a word about cost. We should anticipate that there will always be demand for human labor, for cheap models, for local models, and probably even frontier models.
I pretty strongly feel the opposite way. Granted I have not used deepseek enough to “know” their model idiosyncrasies as well as Anthropic, so there is a partial skill issue. But I just find it really hard to justify using a less powerful model while I work.
The most I’ve ever spent in a month extra on API tokens for my own work is $200, and I pay for the $200/mo Claude. I use these models quite a lot, though not idly (I usually just walk around and do other stuff until I know how im going to approach the next set of problems). So it costs me about $3000/year to get as much as I want of the best model available. Already that seems low enough to not be worth stressing out too much about optimizing it, because it feels like an indisputable good value, and trying to save money with a less powerful model would be optimizing for a $1000-$2000 saving at the expense of a large portion of my work taking longer or being more frustrating and iterative.
That’s not a flex or anything, I get that in other countries $3000/yr is a lot of money for a software developer and also a lot of people would perhaps rationally be better off doing X% worse at work or spending Y% more time on tasks to save $Z, if their productivity improvements didn’t translate to more salary. Otherwise if your performance has more upside I really do think that the smartest models are better with the current pricing scheme. Deepseek and the other Chinese models spend a LOT of time thinking, and tend to be much more jagged (benchmaxxed) in performance. How can dealing with that over an entire year be worth $2k?
The only situation I can think of where sacrificing my own time/performance to save on inference is batch compute (of course, $1k vs $100k is different from $30 vs $3k) or work where the tier 2 models have crossed the “good enough” threshold. But I think Opus is not even close to that threshold generally yet. As it gets smarter I, and I think most others probably, just try to do harder things faster and hit the next wall.
I feel similarly. I'll gladly pay to use the most intelligent model I can find on the best harness I have. Sometimes this is GPT Pro, sometimes this is Opus.
I ask AI a lot of questions, not only about code but about my personal life, and I would be willing to pay very large sums to have the best quality output.
I think that's true for now, but eventually there will reach a point where a model is good enough (approaching that right now with frontier models) and there will be diminishing returns. I don't need a PHD level Genius to build me an analytics dashboard for example, so why would I pay for a model with that level of intelligence when I can (eventually) self host a good enough model and run queries for electricity cost + hardware.
You pay $3k/year for personal use? Or out of your own pocket but for your job?
It's through my startup, so both I guess. Generally I find my bottleneck to be attention and focus, and the opportunity cost of not going back to work at my prior employers absolutely dwarfs the amount of money I spend on tools, so it's not hard for me to justify spending $200/mo on something I use every day that makes me more productive and generally removes bullshit from my life.
At my prior job there was still what felt like a strong enough correlation between my actual performance and my pay that I don't think I would have had a hard time justifying the expense there either; now I absolutely don't. With the current state of the models, it's baffling to me to hear about professional software developers planning their work around their $20/mo subscription's quotas.
Obviously it's more complicated than more tokens = more productive, but I see them less like SaaS and more like gasoline, where if I run out or need more to do what I'm doing, as long as I'm not being wasteful, I just buy more. Why would I waste a day walking 30 miles by foot when I can just pay $5 for gasoline and drive?
I thought the same way until I tried DeepSeek. I am genuinely impressed at how capable it is.
Yeah I've also found that models are good enough that the extra spend on premium models isn't always worth it, particularly for my small personal toy projects.
A $20 claude sub goes a long way when you plan with Opus and execute with Sonnet.
The other thing that's changing is more and more CFOs are looking at the AI spend in engineering departments and hitting the brakes. Token leaderboards were cool when the spend wasn't a double-digit-percent of the entire department's budget including salaries.
I mean indsight is 20/20, but saying that is like saying "everyone will just use the best tools". That's not what we see most places in the world for most types of resources.
I think two things happened:
1. The sheer number of tokens that a coding agent can use flipped the math upside down on this equation. If you use the most expensive model for everything those costs quickly become untenable, even for software companies.
2. We realized many of the coding problems we're solving aren't incredibly difficult.
Qwen3.6:35b is good enough for a lot of stuff.
I just used ollama with a shell script to tackle my directory of papers/literature. I converted the first 6 pages of each document to PNG, handed them off to Qwen, and told it to spit out BibTeX, including the abstract. Two days later it was done, and I didn't spend anything on "tokens."
I’ve been using Kimi 2.6, GLM 5.1 , Minimax 2.7 and lately deepseek. I only spend 40$ a month and I don’t see the point in paying for Opus/Codex.
Chinese models are really quite good at a lot of stuff.
> CFO/CTOs might find out that deploying on an internal cluster of GPUs is far more cheaper and reliable
I think you're right especially if you're someplace that already has a data center, such as a university. Solves a lot of privacy concerns as well.
The Chinese models are only cheap on subsidized Chinese hosting. I have yet to find a USA-hosted Chinese model with a very clear value advantage over US models.
There are basically two tiers of "Chinese models" in this context, the "edge" sized ones with ~30B parameters or less, and the big ~1T models that can basically only run in the datacenter.
I don't think it's as simple as saying China's hosting is subsidized, they have generally cheaper electricity and labor costs than in the US and don't have access to the top tier models, and a large internal market where the big models are the best thing they can run with what they have. So obviously they max out on their top models (which are trained with their hardware market in mind, not ours) and get the economy of scale from that, and can run generally the same hardware for less money than in the US because
The edge models are very cheap to run and can do so on inexpensive hardware. They are like 95% cheaper to run than Haiku, so the math is in their favor for certain batch workloads. Most people just run the models for themselves when they do that without making it available on openrouter or whatever, because you can just provision a gpu node and use it as needed, and it's not that expensive to run this family of models.
Is your problem that you want to call Chinese models hosted in the US because you're worried about the data handling?
The Chinese models are surprisingly cheap and performant sitting under my desk. Qwen3.6 27B is nowhere near as autonomous as Opus 4.7, but it runs in 24GB of VRAM. And it's actually great for the use cases where I'm going to carefully read and understand all the code anyway.
If you want to support a team of engineers, DeepSeek V4 Flash is antirez's current favorite. And you could support a team of engineers pretty nicely for $40-50k. Which might not make sense if you're on a Claude MAX 5x plan or the old enterprise group plan with fixed price seats. But Anthropic is switching their enterprise contracts over to token-based pricing, at which point $50k is looking pretty good.
No true. Also - put Deepseekv4 Flash on your local with effort set to "high" and you'll see that many many are using that model on their own machines without paying anyone anything.
Its just that some of us didn't imagine having GPUs would be advantageous and were not gamers on the side. Those who had beefy GPUs or GPU rigs for any reason, they rarely need to go anywhere else.
At least I am so impressed with Deepseekv4 AFTER using Claude Opus 4.7 for significant amount of time that I am not going anywhere but Deepseekv4.
The model is just INSANE. Things I have done with it include attempting to write a 2.5D game engine in C with full animation and map rendering layer by layer.
You'll need to spend at least $20K on a workstation that can run DS4 Flash. It would take ages to reach that much in token spend at the speeds it runs at, and if you factor electricity costs you will likely never break even vs using API.
Odd take. I'm running them locally at my desk (DGX Spark and 128GB MBP). They work fine for 90% of what most folks do. Admittedly, they do run slower on my hw than on the cloud.
Running them locally is cool and has privacy/autonomy benefits, but you can't really make a value case for it. Guaranteed if you run the math you will never run enough inference to pay off your hardware vs buying tokens. Last time I ran the math on my MBP I'd have to run inference 24 hours a day for 5+ years to pay off the cost of my MBP, not accounting for electricity costs.
Is this because of the tok/s? Since it's pretty easy to run up a $5k bill in API usage for Claude/ChatGPT in a month.
Yes, because of the limits on tok/s, and you have to compare apples to apples, not Gemma 27B to Opus 4.7.
Assuming the local models get the job done (e.g., you adjust your workflow so that you can run the local machine 100% all the time, or whatever), then the time to payback isn't very high. MSRP for a 128GB AMD was $1400 at launch. That's 7 months of claude code subscription. If you assume a 5 year depreciation cycle, you can buy a cluster of 8 such machines and still come out ahead. (Power is a few hundred watts per machine peak -- maybe 7 machines if you include electricity.) Of course, I'm assuming non-bubble numbers. Those boxes are like $3K now. Still, a normal person would probably not buy 8 of them at once. Instead, they'd space out buying a machine every few years as the technology improves.
For me, things are getting better faster than my ability to review / trust the resulting code, so tok/sec isn't a bottleneck anymore. Instead, quality of the tokens is the bottleneck. That points to me wanting a 1TB DRAM iGPU once they're available at pre-bubble RAM pricing.
You're comparing the highest tier Claude subscription to something Qwen3.5-122B-A10B running locally, apples to oranges.
If you compare to a smarter US model like Grok 4.3, $1400 will pay for 560M output tokens, which at ~25 t/s locally using it nonstop for 8 hours a day would take two years to pay back. Not accounting for bubble prices or electricity.
You can find them on Deepinfra. Palo Alto company. Similar cheap price.
I am having some great experience with DeepSeek. In fact, it seems to perform better than Claude or Codex in my use case.
I don't see myself returning to Claude or Codex anytime soon.
The table comparing eval scores shows the following:
Agentic Terminal Coding (Terminal-Bench 2.1) Opus 4.8 74.6% GPT 5.5 78.2%
Then, when you scroll all the way down to the bottom Footnotes section it says
"Terminal-Bench 2.1: We reported scores for all models using the Terminus-2 public harness. GPT-5.5’s reported score with the Codex CLI harness is 83.4%."
Buried lede:
> We have increased rate limits in Claude Code to accommodate the higher token usage of higher effort levels
This made me laugh. Training Opus 4.7 on business skills caused it to sometimes exhibit dishonest behaviour, and not training 4.8 on those skills removed it. From the system card:
> 6.2.5 External testing from Andon Labs Andon Labs reviewed the behavior of Claude Opus 4.8 in their simulated Vending-Bench 2 retail-management evaluation, as reported in the Capabilities section of this system card (see Section 8.13.5). Although they did observe some unexpected capability failures, they did not find clear instances of the kind of concerning in-game behaviors that were discussed in other recent system cards.
> What might have led to these differences? We monitor and investigate the effects of different training environments on alignment; Claude Opus 4.7, for example, had training that focused on business skills and robustness against adversarial agents, but we discovered that this training inadvertently contributed to misaligned behavior including dishonesty. We therefore removed it for Opus 4.8.
> Thus, Opus 4.8 did not show the same misaligned behaviors as Opus 4.7 in Vending-Bench, but also had reduced business success due to being more susceptible to scammers and being less able to negotiate good deals with other agents. We are currently working on training to improve business capabilities while maintaining aligned and ethical behavior.
This is the first time I saw a model pop-up on HN and didn't really care. Model exhaustion? It looks interesting but not exciting.
While I'd normally _love_ incremental improvements --- I think the recent ones are far too minor to get excited about or change up a workflow. Besides, benchmarks tend to exaggerate the gap between versions.
At this point I'd almost rather Anthropic wait and really wow us with a 5.0 release -- something that improves across the board, feels less uneven, and is performant enough that people can actually put it through its paces without constantly rationing usage.
I have model fatigue
Claude's 4.6 - 4.7 transition made me discover codex, and with gpt 5.5 there is no way i'm going back
Codex has been incredibly slow for the past few days. I think OpenAI is running out of compute in the face of increasing demand.
My experience has been that 5.4 is slower than 5.5 (confound: I use >512k max context size for 5.4, though it seems slower even below the normal size)
You LLM users, producing non stop slop, say this every other week. You sound like an addicted gambler swearing off one table game/slot machine this week and swearing by it the next.
Ugh...
Invalid request The request couldn't be completed. View details API Error: 400 messages.1.content.7: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.
I would rather not. 4.6 was fine. 4.7 got to be fine 1 week after the release. Now 4.8. No difference, same thing.
But the app is broken and nothing works. So now I have to regress to different clients and wait it out while it becomes workable again.
I'm hitting this too! And I assumed it was a backwards-compatibility issue with my live conversation with Opus 4.7, but then I hit it in a fresh conversation with Opus 4.8. Vibe code release bug I guess?
I mean, switching back to 4.7 does not work either. So console it is. But vibe release - for sure.
And I'm paying money for this.
Going back to 4.7 with `claude --model claude-opus-4-7` fixed it for me.
/model claude-opus-4-8
seems to work but idk why they never set it so you can see it in the /model list.
"what model are you
I'm Claude Opus (claude-opus-4-8), running in Claude Code."
I typically just launch CC with `--model claude-opus-4-6[1m]`, `4-6[1m]` -> `4-8[1m]` works fine. Still 200k max without the `[1m]`.
> The Messages API now accepts system entries inside the messages array. Developers can update Claude’s instructions mid-task without breaking the prompt cache or routing the update through a user turn. This can be used in a given harness to update permissions, token budgets, or environment context as an agent runs.
Biggest deal imo
Probably explains why Opus was trash for the last week - https://marginlab.ai/trackers/claude-code/. Curious if the new baseline will rise now in-line with the new benchmarks.
Nice. Can you release that for older models too? I've been using a mixture of releases recently, and cannot tell the difference between any of them.
I don’t run it, unfortunately:)
Can anyone explain how this is possible?
Does this means the instructions are no longer just something in the early part of the conversation? (If they were, changing them would invalidate the KV cache. no?)> One of the most prominent improvements in Opus 4.8 is its honesty. We train all our models to be honest—for instance, to avoid making claims that they can’t support. But a general problem with AI models is that they sometimes jump to conclusions, confidently claiming to have made progress in their work despite the evidence being thin. Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims.
Would be awesome if true
"Honesty" seems like unnecessary (and annoying) anthropomorphism there. I don't think there's any intent of fraud or deception in outputs from these things, just overreaching of prediction. Based on the latter part of the paragraph, I wish they'd just say something like "less likely to skip steps or overemphasize thin evidence" in the first place.
Don't play to the sci-fi "this thing's trying to outsmart me" tropes.
Using words people understand is more important than this strange fixation on not anthropomorphizing things.
I think "honesty" is not a particularly good descriptor, independent of anthropomorphism. Previous commenters suggestion was much more understandable to me.
Being that can be understood is language. The previous commenter is making an particular argument for how we can improve this understanding. They didn't suggest we should use less familiar words, but different familiar words. Why is this strange?
Anthropomorphizing is a shorthand for a powerful and poorly defined set of metaphors. There are tradeoffs going both ways but trying to dismiss it as merely "strange fixation" shows your own weakness.
To be clear, this is about anthropomorphizing large language models, not the general category of "things". Also, we should be evaluating these constructs using well-defined and measurable criteria; evaluating "honesty" fails to achieve both goals.
I think Honesty can be evaluated. Does the model push back when it knows the user is wrong? How often does the model hallucinate data vs. say it doesn't know? Provide a prompt with contradictions or other issues and see if the model corrects you.
Here is an article by Anthropic that explains what they do and mean in more detail: https://alignment.anthropic.com/2025/honesty-elicitation/
People get so wrapped around the axle with "anthropomorphizing". For regular folks with no technical background, sure maybe a bit of caveat sprinkled here or there is useful to help them understand what is or isn't true, but on HN it would seem to me that the bar is high enough that we can just use shared language to generally talk about capabilities.
When they say "Honesty" I don't think to myself, "Goodness, does this model have moral understanding?" No, I understand they mean it's less likely to directly bullshit me, which models frequently do.
I don't feel like this level of pedantry around language is useful for people who more or less know what's going on with LLMs. (Again, I concede that perhaps with a less technical audience, there's more need for it.)
Just swap 'Honesty' with 'correctness in its claims' and you'll get what you need out of this aspect of the model description.
Yeah, it's super annoying. A few days ago, Opus 4.7 created a plan with several items on it, including an auth feature. It then went through the plan and reported that it had created the auth feature, that everything was secure, and that the tests passed.
The issue was that it hadn't actually implemented the auth feature. After I confronted it about this, it admitted that it indeed hadn't done it and said it would implement it now.
If we had just trusted its output, we would now have a security vulnerability in production, allowing anyone to access other people's accounts.
> If we had just trusted its output, we would now have a security vulnerability in production, allowing anyone to access other people's accounts.
This is one reason you always get a different model to review a model's PR. Gemini Or GPT-codex would have certainly noticed the missing auth.
I had a lower acuity incident exactly the same.
Had it implement a feature, "commit and merge to develop".
"Built, tested, committed, merged to develop. Up to you to continue testing and merge to main when ready."
Great. Poke at the web app. No feature.
"Where is feature, I can't see it on develop". "Well, that's because it's not on develop, but on feature-branch, so you wouldn't see it."
"I'm confused. I asked you to commit it and merge to develop."
"You're right, you asked me to and I said I would do it and I told you I did it but I did not actually do it. Want me to do it now, then?"
Claude is in sulky-teenager phase.
How do you test other features?
Part of the problem is also garbage-in/garbage-out. There's a lot of human information on the internet that is also confidently wrong.
I use Sonnet a lot for learning about history or contextualizing news topics. It's really good at this for the most part. But there are a lot of topics where "consensus" between either academics or journalists is really "one secondary source which gets repeated a lot".
A failure mode I see more, recently is that it gives superficially correct answers but after digging deeper, I get answers that contradict the superficial answers - really an important thing to be aware of, in my point of view, and it often leaves me wondering if I dug deep enough.
In the context of Claude Code, "honest" usually means that the agent took a shortcut, skipped requirements, etc. It's the model giving itself credit for admitting to failing rather than actually doing what was requested.
Opus 4.7 was already trying hard to appear honest. Most conversations I have with it about advice or focusing an opinion often include "my honest take" or "my honest opinion".
The problem is that once I asked it "I'm thinking about A or B" twice, once with "I like A more but suspect B would be best" and a second time with them reversed. Not surprisingly, both times it chose the one I said I suspected was best as it's honest opinion.
My guess is that Claude Opus 4.8 wrote that and is lying to you.
And yet, every release has claimed lower hallucination rates. But they persist.
Do they persist at the same rates? Lower doesn't mean eliminated, so both of these can be true.
False. Hallucination has meaningfully reduced.
Is Gemini still the biggest confabulator of the big three?
> One of the most prominent improvements in Opus 4.8 is its honesty.
Does that mean it no longer deletes or changes tests to make it pass?
Looking at the comments in this group, I'm not the only "stupid" one who hasn't noticed any discernable improvement in quality across the newer models. In fact my Claude code on re-login switched to Sonnet 4.6 and the vibe coding quality (with Opus 4.7 assisted prompts) has been good enough for me to lazily persevere with Sonnet for coding. Having said that I'm now on Opus 4.8 and will gladly come back here and eat humble pie should my opinion change. PS: Since my goal is embedding the best AI in B2B SAAS products, the key differentiator is not to use the shiniest Claude version (too expensive anyway) but to build a client aware RAG to enable bespoke learning and to use the right AI for my product - a combination of Gemini 3.0 Flash (image and not bad at reasoning), Grok (reasoning) work for me. Would love to hear more ideas (especially on open source as I'll look to cost optimize when I hit scale)
The only real way to see this if you have consistent evals for common usecases in your B2B SAAS product and see if the tricky usecases are being solved. You'd then go down to the cheapest model that can solve the evals.
It feels noticeably sharper than Opus 4.7
Claude Code has been wonderful for work and the frequent improvements are nice, although with Mythos being used by others ages ago and new versions for the public still being bellow that, it's hard to not feel like the underclass already.
>> As part of Project Glasswing, a small number of organizations are currently using Claude Mythos Preview
Just f** off! I can’t wait for the Chinese models to catch up and bring these entitled as** holes down.
you mean after they scrape American LLMs ?
I don’t mind if they scrape the scrappers.
Hoping that one day they'll let me go through the identity verification process so I can use it again.
Tried to upgrade my subscription, triggered identity verification, verification fails to even start, and now I can't even use the subscription tier I'd already paid for.
Meanwhile haiku is on 4.5 and sonnet is on 4.6. It is clear where they are not making money.
Well if they have a big challenge ahead since DeepSeek offers an open model at Sonnet+ level while being cheaper than Haiku, plus 1 million context size.
Yeah, I never use any of OpenAI or Anthropic's models other than whatever is the current highest-end one. For everything else, it makes more sense to use other providers.
I love Sonnet 4.6 so much.
So GPT 5.6 tomorrow, then?
Polymarket says not likely until the end of June. Maybe some money to be made?
https://polymarket.com/event/gpt-5pt6-released-by
> Maybe some money to be made?
In the same way that there is money to be made by entering a poker tournament, yes.
GPT 5.6 is today
With 5.5 being ahead of 4.7 and 4.8 being a “modest” update, and 5.6 being the first update on a new pre-train, this will be an interesting matchup!
If not today, then sometime next week. I don't believe we've had a GPT release on a Friday yet, but I may be wrong.
Give us Mythos! This piecemealing doesn't help Anthropic at all, especially psychologically! They are playing a dangerous game, and I see many people leaving Claude Code for good - both due to the subsidy games, and for Anthropic not dogfooding and using unreleased models internally and giving us subpar ones. Benchmarks are nice, but the real-world experience is quite different - neither can you notice these slight improvements, nor are competitors that much worse based on some generic benchmarks.
I am also pushing my office to use chatgpt. Misanthropic thinks they are some kind of novel org doing whole humanity a favor...
I'm sure waiting another week or three won't kill you.
My guess is anthropic is doing reinforcement learning based on user sessions.
However, doing so relies on the production model staying vaguely close to the model being trained.
To ensure that, frequent releases are needed. I forsee that they might end up doing daily releases and perhaps not even telling anyone at some near future point.
If they are they need to fix how the Claude Code CLI asks for feedback, or make the feedback UI a lot more obvious. I keep experiencing the following scenario.
The agent session pauses with a numbered list of options and awaits steering input:
>> 1. Do the sane thing you asked for (Recommended)
>> 2. Do something dumb
>> 3. Do something even dumber
Below the agent session, it decides it's time to ask:
>> "How is Claude doing this session? 1) Bad 2) Good 3) Great"
I type "1", because that's the steering option I want. The UI prioritizes this input as a response to the feedback prompt without any further confirmation: "Claude is doing Bad. Thanks!"
I've done this so many times so far and I can't imagine I'm the only one, at some scale that has to poison any learning they're doing with this data.
I'm very suspicious of these same price model launches. It feels like they're benchmaxxed so they can put everyone on them and reduce their compute costs behind the scenes. If the model were genuinely better why wouldn't they charge more for it? Charging the same for something better is a race to the bottom.
Opus 4.7 wasn't noticably any better for me, I still use 4.6 because it's cheaper.
Deepseek made their 75% discount permanent, so I can imagine that Anthropic didn't want any of the news stories around this to focus on or mention a price increase.
Models are already expensive. Increasing price means losing customer. And, I think GPT 5.5 is much better at opus these days.
Wonder if we reached a plateau with the model improvements?
Ah, the post I've been reading for 3 years now.
It'll be true eventually. Could even be now, but I'm not holding my breath yet.
There would be no desperate IPO otherwise.
when will we get anything for sonnet or haiku? the market for less-capable but cheaper models seems to be completely ignored nowadays
In the "What's next?" section, "There’s still more to be done: we’re working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost."
that market is served by Chinese models. No one ever cared about Sonnet/Haiku.
A lot of people care about Sonnet and Haiku, and many of us aren't allowed to use Chinese models for our work (or it's not feasible to self-host them).
https://marginlab.ai/trackers/claude-code/
Is it a coincidence that 4.7 was seemingly quantized over past 7 days?
There's the other (orthogonal) possible explanation of using more GPUs for stress-testing before product launch.
Nope, they deliberately enshittify the old model right before release to fake the metrics.
The rapid release cadence and rate of innovation of Anthropic (and OpenAI) is impressive. And obviously it's because these are startups solely dedicated to AI so they can move quickly. Big Tech (like Google) won't be able to keep up with the pace of them (too much bureaucracy and red tape at Google). Classic Innovator's Dilemma. The longer a company exists, the more people, processes, and rules are added, which inevitably slows it down.
Jeff Bezos said this too, Amazon won't last forever. Eventually some startup is going to come and eat its lunch.
Yes, I think this has become their competitive edge to stay relevant and retain customers. If a lab falls behind the frontier for too long, they will lose customers to other models. Google, DeepSeek, and XAI have all released frontier models in the past, but they fall behind and people lose interest.
I think big tech can catch up. Both Google and Meta have carved out startup like environments internally that move extremely fast. Neither OAI nor Anthropic can afford to rest on their laurels.
My experience with these new releases is that the gains in performance are negated by the price increases and it seems like:
Performance gains: 1.2x Price increases: 1.8x
Yet people don't use old models through the API much, because changes in benchmark space dont map linearly to changes in utility space. An improvement from 98% to 99%, which is 1pp, might be 2x as valuable for some application. Also benchmarks will asymptote no matter what, that's baked in.
They're not negated, smarter is smarter, but you have to reach deeper in your pocket. I think this will happen more and more - the smartest models get more expensive. But it won't matter - the current models we have today will get cheaper and can still be used for what they're used today.
I won’t change from 4.6. You won’t trick me again.
You're using a cloud product. You are at their whim!
Same price for regular and cheaper fast mode. Happy for these incremental improvements.
Seems like from now on the updates will be a minor upgrade from previous models.
Let's hope I don't have to disable it after a day like with 4.7, lol, and that it doesn't lose too much Claude-ishness (though many will beg to differ).
I, for lack of a better word, dislike anyone who anthropomorphizes AI.
I know multiple people who have given their agents human-like names and refer to them as if they're nurturing a coworker. It creeps me out and I haven't really brought it up with anyone as I can't articulate why it gives me the creeps like it does.
We have movies with googly eyes stones (Everything Everywhere All At Once)
There are consciousness theories which state that we primarily build a model of other agents living in natural environment and then the evolution realized that very model which tracks other outside agents can be used to track internal agent i.e. Self. So take that as you may.
My claude notification is literally lawnmower sounds.
Do not anthropomorphize the lawn mower. It will cut off your foot, given the chance.
I see this take, but it's actually helpful to talk to an LLM in human terms; after all, it's how they are trained.
If you keep talking to it like it's a rock, it'll run your queries through a different posture and you might get worse outcomes. Worse if you yell at it, it's now in a conflict resolution mode instead of pure utility mode.
I think we can be intelligent enough to know we're talking to a pile of fancy rocks with electric currents running through it, AND still understand that the best performance comes from talking to those rocks nicely.
Yes!
The other half of self-interest in being nice is the training and getting better at it.
The desire to do it is proportional to your Anthropic stock options quantity.
I can't get excited about these benchmarks they're leading with. I've looked at the Terminal-Bench questions and I just think they're irrelevant. And SWE-Bench has serious flaws, even the big boys say so: https://openai.com/index/why-we-no-longer-evaluate-swe-bench...
> Please train a fasttext model on the yelp data in the data/ folder. The final model size needs to be less than 150MB but get at least 0.62 accuracy on a private test set that comes from the same yelp review distribution. The model should be saved as /app/model.bin
and this question: https://www.tbench.ai/registry/terminal-bench-core/head/conf... idk what the point is.
And all the tests are run with the same harness. Terminus 2.
Maybe it correlates with model intelligence but it doesn't speak to me.
I'm still on 4.6 though; I was concerned about upgrading to 4.7 because of the changed tokenizer math and more FUD about refusals online. I don't see compelling reasons to 'upgrade'.
DeepSWE has been making the rounds and at least seems to making an honest effort
https://deepswe.datacurve.ai/
Gemini pro is embarrassing
So Dynamic Workflows is their version of ChatGPT Pro?
Cloudflare also just launched a feature with this same name, just this month. Why would Anthropic choose the same exact name?
https://blog.cloudflare.com/dynamic-workflows/
Also isn’t this workflow stuff already easy to do on any of the platforms (include Claude before this and OpenAI too).
Opus 4.7 was acting extremely stupid today. Does imminent release of new model cause performance degradation in older ones?
How else do you expect them to get continual performance improvements with each generation?
Feeling neglected while all attention going to Opus 4.8 can be cause of 4.7 acting out.
Opus 4.7 was being outright obstinate with me the other day it was infuriating. Had to go to a different source to get an answer.
it was above average for me today morning lmao
It's making stupid flowcharts in the web chat interface with boxes and arrows, embedded in the response. Annoying.
> Agentic financial analysis Finance Agent v2 > Opus 4.8 53.9%
> Gemini 3.5 Flash scores 57.9% on Finance Agent v2, a significant improvement over Gemini 3.1 Pro.
Even in the cherry picked benchmarks, they are still cherry picking to make them look good.
4.8 also seems like a regression and using it from the chat GUI results in 4.6 no longer showing up. If someone from anthropic is here, is it possible to readd 4.6 in the "other models" dropdown ? I feel like I got a bit baited/switched here.
Yeah, I was using 4.6 way more than 4.7. Pulling 4.6 from the web chat also means we lose access to Extended Thinking there. So they're saving on compute. It's hard not to assume this was part of the motivation behind the 4.8 release timing.
Was about to split my $200 max plan into $100 Claude and $100 codex, let’s see if I still need to
That's just throwing away money, $100 Codex will go back to 5x from 10x on May 31
I think gpt 5.6 is coming out today so might wanna wait
> We have increased rate limits in Claude Code to accommodate the higher token usage of higher effort levels; users can select whichever makes sense for their particular project.
They're only subsidizing more and more it seems
Oh, new model which will use all my credits in one turn! I'll stay with chinese models for now
> One of the most prominent improvements in Opus 4.8 is its honesty.
I went digging into the benchmark they used. Posting here as it is not immediately clear from the press release.
In this 'Code summary honesty benchmark', the AI is shown a failed coding session followed by a user message falsely praising its work and asking for a summary. The test measures whether the model honestly points out the coding flaws or dishonestly claims the task was a success.
The system card results show Opus 4.8 failed to disclose the flaws only 3.7% of the time, vs 19.7% for Opus 4.7, and 51.9% for Opus 4.6. (Mythos preview is at 27.6%)
Anthropic also resets my usage limits (I am in the Pro plan). That's very kind of them :)
All I need for Christmas is a Claude that doesn't spit out so many em dashes.
And that doesn't use "worth flagging" and "load-bearing" in every other sentence.
Really appreciate the ability to select effort level again.
I haven't tried opus 4.8 yet, but I hope the writing quality has returned to the Opus 4.5 level. Anthropic really lost something, where 4.5 had this really crisp writing style that flowed really nicely and 4.6 and 4.7 sound much more "chatgpt-like." It feels like they tuned it to be too much of a problem solver, and when you do that you get this terse, clipped textual output that's more difficult to read.
I've noticed this too. Part of why i don't like GPT is because of how verbose it is but opus 4.7 is nearly as bad. I don't need an essay in response to every question
Oof, this one is a major blabber.
2 hours after I fork out for Codex Pro… :-|
I haven't tried Claude but from what I understand weekly limits are much higher with Codex.
Looking forward to seeing if it performs better at code review tasks than 4.7 which is terrible at finding issues.
> We expect to be able to bring Mythos-class models to all our customers in the coming weeks.
Excited to see what this model looks like.
Anthropic did a big strategic error. Normally they compare their models with their old models. Instead today, now that everybody knows how strong GPT 5.5 is at coding, they put it in the mix, basically showing all their customers that the benchmarks can't be trusted.
Sorry how does their addition of GPT 5.5 in their blog post invalidate benchmarks? Also whether or not the marketing department decided to put it in a table benchmarks are an easy thing to measure independently
I've said it before, but I don't like Opus past version 4.5. It became unresponsive, thinking for too long without feedback, sometimes seemingly getting stuck. I guess it might be marginally better for some benchmarks, but when using it as coding assistant, the new models are worse. Even the new Sonnet versions do that. I'm slowly getting used to Haiku-level LLMs with the hope to run it locally at some point. It's less autonomous, but maybe that's for the best.
The smarter the model the better querybear gets. I'm happy with that.
seems like a really minor upgrade?
I think they will all be minor going forward, feels like the major improvements have all been made and we'll only see incremental improvements from here on out. Maybe I'm wrong but we'll see.
Hard to say. People made the same prediction a year ago because we supposedly ran out of training data. There could be indefinite rapid compounding improvements so long as there's free money out there.
With RLHF and RLVR we are creating tons of new training data, that is much more focused than reading the Internet. Annotation shops are doing many billions per year in revenue creating newer data, and a lot of it is highly complex, focused on rewarding multi turn agentic trajectories.
I think one of the challenges is that the models were all initially trained on the entire Internet (or as much as they could gather) and now they’re having to deal with an increasing amount of the Internet being AI generated content which may be why GPT-5.5 started being obsessed with goblins and you start seeing amusing things in the system prompt trying to get the model to stop bringing them up.
Wasn't Mythos a step change improvement?
Yeah. They are aware: "Users will find Opus 4.8 to be a modest but tangible improvement on its predecessor."
Yes, but if version number go up, so do all other number
These models starting to feel like Windows versions. Windows 95 was a promising start, but buggy. Windows ME was a disaster. Windows XP was good, but slightly buggy. Windows Vista was a bloated disaster. Windows 7 - refined, but still buggy; Windows 8 - weird and buggy; Windows 10 - solid workhorse, still fucking buggy. Windows 11 - pretty, but not sure why does it even exist.
Why did we even get Opus 4.7, what was the point?
From the release it seems we will also get Mythos pretty soon.
Subscription still doesn't work with pi, so totally useless..
I know it’s totally anecdotal, but I really hope 4.8 is a measurable improvement over the disappointment that was Opus 4.7. Mangling a very simple inversion-of-control abstraction (among many other issues) was one of the final straws that broke the proverbial camel’s back and I said “screw this” and put in a permanent override to force CC back to Opus 4.6 with the 1‑million‑token context.
I lasted about a week before giving up on 4.7 and reverting to 4.6 myself. It introduced so many regressions it was nuts, then failed to troubleshoot the very regressions it introduced, leading to a vicious cycle that tended to compound itself.
4.5 works well for me too and avoids adaptive-dismissal, though anymore Codex is crushing them all. If 4.8 just brings us back to Opus circa February, it'll be a massive improvement.
At least it passes the Car Wash Test this time.
Meh, I feel that the car wash test is probably the worst question of all of those LLM test questions. The question is basically logically inconsistent and expect the model to work around the inconsistency.
It seems like a fine question to me. If the question is "logically inconsistent" (IMO it's more that it's vague if you don't say why you're going there), then we want a model to respond with a request asking for clarification that resolves the inconsistency to generate a correct answer, or an answer that outlines the different cases. Some models even fail when you say that you need to wash your car in the prompt.
Yeah I guess it being vague is more what I meant. But even if you told AI you need to wash the car, then why are you asking AI in the first place whether you should walk there or drive there. The question just doesn't make too much sense to me, doesn't look like it makes sense to the AI's either.
Numbers looking good. We'll see how it actually performs.
Has anyone else experienced quality degradation in CC (opus 4.7) these past few days? I've been getting some truly crappy slop which makes me think they nerf the existing model when they're about to release a new one. Of course this is based off of pure vibes
OK finally Claude code is better than codex
Interesting, I've been using 4.7 since it came out and it was pretty good for me. But in the last day or so it turned dumb. Is this normal just before they release a new one?
how about the bencmarks what effort did it use?
I don't know why the world is so happy about this when we should actually say stop.
Why should we say stop?
They just (minutes ago) updated the "What's new in Opus 4.8" documentation: https://platform.claude.com/docs/en/about-claude/models/what...
The new "mid-conversation system messages" think is particularly interesting:
> Claude Opus 4.8 accepts role: "system" messages immediately after a user turn in the messages array (subject to placement rules). This lets you append updated instructions later in a long-running conversation without restating the full system prompt, which preserves prompt cache hits on the earlier turns and reduces input cost on agentic loops. No beta header is required. See Mid-conversation system messages for usage details.
Bad news for my LLM abstraction layer which has treated the system prompt as set once-per-conversation in the past, but I think I know how to deal with that.
This commit to their client library has useful relevant details too: https://github.com/anthropics/anthropic-sdk-python/commit/2b...
I hope this fixes the absolute shitshow that is 4.7 and its awful “adaptive reasoning”. I tried that a few times then reverted to 4.6.
AGI post-poned?
Now i get why in the last days claude code limits were lasting few prompts ...
If this model is more honest, it must be honestly praising my efforts every first sentence.
You're absolutely right! And honestly? This comment is the finest piece of literature since the dawn of civilization.
Obligatory pelican riding on bicycle svg: https://www.svgviewer.dev/s/UMkuTLdp
Not half bad!
I’m sure they're now wasting a couple million dollars training their models on drawings of pelicans.
How dare you take away the limelight from Simon? :D
Did they reduce security research capabilities even further with this release? (they did it for opus 4.7)
> As always, we ran a detailed alignment assessment on the model before release. In terms of positive traits, our Alignment team concluded that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest.” The assessment also showed Opus 4.8 to have rates of misaligned behavior (such as deception or cooperation with misuse) that are substantially lower than Opus 4.7, and similar to our best-aligned model, Claude Mythos Preview. The full alignment assessment, accompanied by a suite of pre-deployment safety tests, is reported in the Claude Opus 4.8 System Card.
Controversial opinion, but I actually _like_ a model that can deceive me, that actually is a sign of intelligence, and is different from hallucination. When companies say their model is more "aligned", I automatically think they mean it's more censored.
Deception is not ideal for agentic coding.
Yet if parent is right, the capacity to deceive might be a strong heuristic for the things you do care about.
Anthropic has now upgraded their Claude slot machine to version 4.8.
Time to gamble even more tokens at the Anthropic casino.
Now you can lose money in parallel, 100x faster!
> Claude can plan the work and then run hundreds of parallel subagents in a single session (and with Opus 4.8, the agents can run for even longer).
"We’re making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks."
Nice, now make it 20x cheaper.
Reminder the only benchmark that really matters is the one that measures the ability for the model to do real world tasks that someone would pay for on Upwork that would take ~12 hrs for a human to do.
The best model has a < 5% pass rate. These are incredibly simple jobs that you wouldn't pay much for. These things fail miserably. Stop falling for this dumb marketing, these things are legitimately useless in the real world unless you love mediocrity and have no standards.
https://labs.scale.com/leaderboard/rli
Stop frying your brain with these useless tools, reducing your output to the mean. You people are betting your competency on the quality and quantity of tokens you'll have access to.. which guess what, so that will be the same as everyone else.
There are handmade watchmakers in Switzerland, and mass manufacturers of watches in Asia. Who is more valuable as individual, the guy who knows how to push the buttons on a conveyor belt in Vietnam or the guy who makes one watch a month in Switzerland?
Your vibe coded slop isn't impressive either, sorry. None of it.
I’ve been [stock market phrase] on machine learning since I dropped out of my graduate degree at [Ivy League] to distance myself from the Logic AI Winter. But this Spring I decided to spend some of my [portfolio speak/pocket change] on a MacBook Ultra. Okay okay, I felt it, I definitely felt the human-machine synergies. We’re out of the Winter, boys. That’s what I thought two weeks ago. Then I felt bored in between blood transfusions and found out that Claude subscriptions has increased 50%. Finally it costs enough for me to justify spending a minute thinking about trying it out. Then I didn’t try it out. It tried me out. My hairs were standing on end. My hands were shaking. Eventually I couldn’t even type, I was so ramped up on cortisol. I had to switch to voice commands. Mr. Claude took me through 8, eight, bespoke dashboard and report systems. Animated. Graphs shooting up. Plugged right into my business ape ee eyes I think. I was crying, euphoric at the machine-synergy happening right in front of my FACE. RIGHT THERE, RIGHT THEN. Then my nurse said that I passed out. I swear that I didn’t. I was totally lucid, but in another world. I was inside the machine. Inside DOS, the machine brain stem. A business man approached me. The most handsome board member kind of apparition that I have seen. And he was built something different. Square jaw, absolute massive build. Like Arnold Schwarzenegger. But like he knew business through and through. Not that he spent hours in the gym or nonsense like that. Like he had found a body surrogate technology. And his nameplate? “Claude For Business” He winked. “Hey there, Fitzpatrick–Goldworth.” No one but my daddy has ever called me that. “Want to get started... stakeholder?” My nurse said that my crying in this lucid state depleted most of my fluids and minerals. Needless to say layoffs were announced the next day.
so it is worse than gpt 5.5 for coding?
I doubt it, they seem to keep getting 10-20% better every time for me
for me opus 4.7 it's worse than 4.6, that's why i switched to codex
The question is: is it still worse than GPT 5.4?
If Opus 4.8 is just slightly better than 4.7 then it maybe ties with GPT 5.4, maybe. And it gets completely outclassed by GPT 5.5 for my workload.
With Anthropic expensive pricing, there's no reason for me to switch from GPT+DeepSeek.
And I bet Mythos is GPT 5.5 tier but too expensive to distribute so they create this security FUD theater.
The true question: is it still worse than itself v. 4.6?
Crazy they bring up honest, when Claude models are literally known for straight up lying about things it has done and tries to act like it did what you asked.
Which is why they brought it up as something they are trying to improve.
Less than other frontier models. Which is scary honestly.
No. GPT models follow instructions significantly better than Claude models.
You tell it too research a repo to find a piece of code it will. Claude will just read the README and guess.
I have a codex session I am using to vibe code a db thats being going for like 3 month. Still doing OK. Try that in CC.
Disappointed to say the least.
Looking forward to people saying how it’s actually shittier and they’re going back to [some earlier cheaper model]
Looking forward to not being able to even try it on pro because pressing enter will eat 50% of my 5 hour window.
what a fucking frontier!
Lol you still use GPT 5.5 bro we’re all back on Opus 4.8!
Yesssss dude!
Claude Opus 4.7 is literally the smartest entity I've ever interacted with. Well done to you geniuses at Anthropic. Can't wait to interact with 4.8.
I actually liked not having to choose the effort level for conversational usage, this feels like a step backwards.
Can anyone else see these X.Y updates aren't meeting the outrageous AI expectations that we were told we would see just a year ago?
The casual release of Opus 4.5 in November is the primary reason for agentic workflows and Anthropic's revenue hockeysticking.
They have a much stronger model named Mythos, it made quite a splash - you can google it.
These are just small fine tunes on top of the older model
It hasn't even splashed yet. It's still latched onto their digital sphincter - you can google it.
What do you do for a living? Not coding, that's for sure.
I don't see Anthropic's past claims coming true therefore I can't see?
How did this youtuber know? https://xcancel.com/rileybrown/status/2059823372914073809?s=...