The fact that AI models can be so easily distilled and replicated is such a stroke of luck.
10 or 15 years ago if one had asked me to envision a future where a private company invents artificial intelligence, I'd have thought for sure they'd have a massive moat, be very difficult to catch, and it would create an almost instant monopoly.
Rather, it seems that selling intelligence might end up as a race to the bottom.
Who woulda thought that just having access to enough textual inputs and outputs and a vaugely similar transformer architecture would be enough to copy-cat rather useful intelligence.
It reminds me conceptually of the idea of using a ST:TNG replicator to just give you another replicator of your own, or asking a stereotypical genie for "infinite wishes". The genie is indeed out of the bottle in many ways.
OpenAI could still have a significant moat. ChatGPT occupies most consumers’ minds when they think about AI and has become a household name. Google won because search became a habit-forming product people grew accustomed to using. Bing was once effectively indistinguishable from Google Search, yet still failed to achieve mass adoption because users had already become accustomed to “Googling” things. The same could be said for people "ChatGPT-ing" things. If OpenAI and Anthropic are smart, they will maintain similar pricing rather than aggressively undercutting each other, allowing the market to resemble Home Depot and Lowe’s, or cloud computing, where AWS, Google Cloud, and Azure coexist as highly profitable competitors. Unfortunately, I doubt OpenAI or Anthropic will pursue this strategy, as both companies appear to be acting as though the race to AGI is winner-take-all even if the market may ultimately support several highly profitable competitors.
> If OpenAI and Anthropic are smart, they will maintain similar pricing rather than aggressively undercutting each other, allowing the market to resemble Home Depot and Lowe’s, or cloud computing, where AWS, Google Cloud, and Azure coexist as highly profitable competitors.
Wouldn't that just be price fixing? If they arrive at their prices independently and they all happen to be similar, fine. But if they're all "smart" and coordinate so none of them undercuts the other, that's probably illegal.
The internet created lots of monopolies with network effects and economies of scale.a low margin commoditized business that still attracted a trillion dollars of investment to get off the ground was not how I envisioned it happening either.
> Rather, it seems that selling intelligence might end up as a race to the bottom.
Personally, I came to this conclusion early this year. To acquire the data that AI Companies are using to train their models is low cost and once they have it, they can refine and store it. Creating the LLM takes a bit of money but it is not a serious blocker. Clearly, the Chinese companies can make AI so they will drive down costs. There is a need for good AI (Not just Great AI) and it is not cost prohibitive to make good AI (The same with specialized AI).
My prediction is that AI will spilt into two categories, Great AI (High Cost) and Good Enough AI (Low Cost). Which for the long run of AI and companies that use AI, this is good.
It reminds me of the seo antics out there. The search results page is the engine, much like how distilling is the "intelligence" for your chinese room machine
Funny you mention Chinese Room and LLMs in the same response, I would say LLMs proved Searle wrong, agents now make cutting edge discoveries and meaningful problem solving. They not lookup tables though and you need to pay for inference, so the intuition of syntax doing the work of semantics without understanding was wrong.
Hmm, don't people think that if the frontier labs really put enough engineering effort into preventing distillation that they would be able to do that, or at least diminish it significantly? I'm sure there are variety of additional techniques they could use on top of what they already do, but I suspect it just hasn't been at the top of their priorities yet. Maybe that will change soon. Worst case they could add additional hurdles to account creation ("know your customer" type of thing).
The frontier labs have competing goals in mind. They want high growth (which means little friction for account creation), API access (because enterprise money is the best money to have), and distillation protection.
Besides, identity verification that actually works at scale is a much harder problem than identity verification which is good enough to satisfy your compliance people and regulators. Especially if the fraudsters have a major world government standing behind them, and if their aim is to be identified as a real customer, not one customer in particular.
Even if it were possible it wouldn't change the outcome. China is capable of training frontier models even without distillation. Distillation is only an accelerant.
The primary resource you need to train LLMs is money and China has plenty of that.
At the end of the day, while you can do your best to obfuscate your reasoning tokens, it's a losing battle to hide actual user-visible output tokens. The very nature of API offerings is that you can't do KYC on where that API's output is going - there's a rich secondary market that's not going away.
And with the sheer volume of data created from that, coupled with benign-seeming prompts like "plan out your reasoning in a document before implementing" that could never be patched without breaking existing customer workflows... there's more than enough for someone to distill on. Even if that only gets them to not-quite-frontier, if you're pushing the frontier every few months, they're only ever a few months behind you.
I took GP as asking for evidence that the reduced-cost Sol is actually a distillation of the previous-cost Sol. AFAIK, providers distilling or quantising models and offering them as the same model have not been proven.
I doubt he was claiming that. He's probably saying that the ability of Chinese companies to be able to distill frontier US models has put downwards pressure on the price of all models.
Alternatively, modern AI is good enough at optimizing its own kernels that it just keeps pushing costs down. Unlike the semi-decentralized inference provider community, OpenAI has both the talent and the compute to throw at the problem of making their models much more efficient to run.
GPU kernel optimization is just the kind of well-bounded problem with clear success criteria that AI loves.
I'm saying that it is unclear that without distillation this wouldn't still be happening. There is a massive narrative that no one but OpenAI, Anthropic, and Google can make a model without distilling. But there's basically no evidence of that.
Even before LLMs, ML folks were already aware that you can use a model to teach another model. I doubt this is something AI companies put at the top of their investor materials, but it's been nice to see it play out.
That said, there are other moat factors like, a US company needing to use a US AI provider, sticky customers due to corporate onboarding friction, and others. Not nothing, but not as large a moat as some imagined.
Yes, but 10 or 15 years ago, I would have thought that there'd be more to it than just a slight modification on the ideas behind a CNN to get this level of AI.
There were somewhat good reasons to think it needed more than just this data-driven ML approach.
There's something startling about how (relatively) simple these networks are and yet how powerful they are. The main ingredient the AI darlings are using is vast amounts of compute and data. I don't want to take away anything from what the researchers came up with, but I suspect even they are surprised at how capable some of these models have become.
early on there was a lot of talk about "emergent behaviors" in the models where they were good at things that were unexpected or did not align to the training data. IIRC doing arithmetic is one example from early on. I think this is where the AGI craze took off, the labs were throwing more and more data in the training to see what other behaviors would emerge. The thought was with enough data and enough parameters AGI would surface on its own.
Then i think tool use became a priority or at lest a sibling priority to more data/more params. Along with multiple specialized models communicating with each other which is sort of a special case of tool use. That pretty much brings us to today.
I think the only moat in the future will be the scale of hardware deployment. If one company is able to deploy an order of magnitude more silicon, they'll have a firm grip on a SOTA model and massive inference usage.
China or SpaceX seem like the 2 likely candidates in 5 years, but who knows.
If (a) demand for AI continues to increase, and (b) SpaceX can get to ~$100/kg to orbit, then they will have a ridiculously deep moat. Probably more like 10 years, though.
They will have moat in the satellite launching business, which is not useful in the AI datacenter market.
You can put AI chips in datacenters in the desert for far less than $100/kg. With lots of solar power available, the option to easily access your hardware and far less radiation issues.
The datacenter in space story really only exists to make it possible for Musk to sell X to SpaceX and make more money from the IPO. That's all. There is no engineering reason.
Musk's bet is dysfunctional politics will make it impossible to build enough data centres and the energy needed to power them. There are many, many reasons that might be wrong. However, if the economics are even close to viable, they could start throwing up data centres quicker than anyone can build them terrestrially (at least in the democratic west).
Yeah, very hard to predict the future at this point. But the Starship + Terrafab combo will be this type of order-of-magnitude-moat IF it works out. Big if.
If it doesn't work out, I think China's exponential terrestrial energy deployment will eventually give them the lead, IF they can get enough chips. Another big if.
well, a stroke of luck until the whole US stock market crashes & everyone's retirement funds get cut 40% I guess when people internalize this. it will have to happen sooner or later though I suppose
interestingly also, open weight models are also more effectively run in the cloud, so it creates a weird scenario where the frontier labs crash but the compute providers, not as much
I wouldn't be so sure about that. The popping of a bubble is usually just as irrational as its rise.
If investors start fleeing from senseless businesses in the AI sector, that does not mean that sensible businesses will be spared. These things follow herd mentality, and the primary drivers of the herd are greed and fear, not fundamentals or business logic.
Yeah the last year has been astonishing, my portfolio is kicking ass. But I'm 10 years out from retirement and I am pretty confident a correction is coming; I hope the correction happens soon.
It’s not as bad as dot.com of course since all purely AI companies are private and the ones on the market have pretty decent cash flow outside of AI. But the stock market pattern is not that dissimilar, the largest increases are usually just before the crash.
The market (s&p500) crashing 40% puts us at levels we haven't seen since 2024, well into the creation of LLMs. Probably a worthwhile trade if it was either/or!
It feels like a slightly more palatable version of what Anthropic has been doing, with their constant "use your free tokens before they expire next week!" campaigns. But it's feeling more and more ominous now, like they've hit the top of the demand curve and need to pull back prices to continue growing.
The top comment on this thread was about AI models being easily distilled being a stroke of luck.
This should not be surprising at all. Every new students spends tiny fractions of time learning knowledge that took many lifetimes to discover. This fundamental to the progress of intelligence and understanding.
It should not be surprising that AI can be distilled. It's the logical method of training; I would hope that each frontier model is in fact not trained 'from scratch' each time.
We should expect future frontier models are simply distilled versions trained by specialist models, the same way humans learn from a series of professors, papers and canonical books on each different subject material. Models like this can be trained incrementally, or a so called Mixture of Experts (MoE).
Once they make a model better than Fable I’ll be switching to Codex. Their priorities in terms of consumers seem to be better. I do think Anthropic has some solid safety viewpoints, but I don’t necessarily think that either is entirely aligned yet with delivering exactly what humanity needs. Maybe the AI will help align the AI companies when it gets smart enough. That’s the real misalignment I’m concerned about.
It feels like 5.6-Sol is already fairly close to Fable, and in some ways exceeds it. Just the other day I had Fable draw up a solution for me, and then I fed it into 5.6-Sol and said how does this look ... it found an oversight and told me about it, and when I then fed that observation back into Claude it acknowledged the miss.
I've noticed also that 5.6-Sol is more concise with output than Fable (and let's not talk about Opus, which is even more wordy).
It's common for different models to find holes in another's work. There are various good reasons for that.
FWIW, we use ChatGPT for our primary model and use Claude to do the reviews. This works better than ChatGPT doing it's own review even with a clean session/context.
Same here. Grok Build 4.6 for me, given how cheap Grok is and how Sol is supposed to be "the" SOTA, it finds a surprising amount of bugs. Most of which Sol agrees with needs to be fixed or improved.
I've done this tens of times between these two models and it works great in my experience. Sol initial back and forth with me. Commit. Let Grok review. Sol fix. Only then do I start reading the code.
Agree, but the point is not because Fable is better than Sol, it's because it's .. different .. it just looks at the problem through a different angle.
I don't think these companies have humanity's needs in mind when they're developing these models. Although the last part of your comment struck me as a bit comical, I genuinely believe that an AI can have way more empathy than a corporation. Afterall, a mimicry of empathy is probably better than no empathy.
It's a funny comparison. Comparing the empathy of some software to the empathy of a company. It's like saying my car was more empathetic than my school. How can those two objects even be compared is what i am wondering
We live in an odd time where 'software' (well neural networks) can be far more empathetic than summed product of a corporation.
Company empathy does exist, just look at how easy or hard it is to reach a company when you have a problem. How do they try to solve it for you? Is it a brick wall, for example Google when you have a problem. People quite often like dealing with small businesses because they can reach a singular human and have them as an interface to the problems they face now and in the future.
Agentic loops and the models underneath them can have a simulacra of empathy too. Not every model just blindly agrees with users, and some have a much better depth in picking up context clues that the user on the other end is having a hard time. Businesses just typically aren't running more expensive and fragile systems like that though.
Well, companies and AI are both entities that can make decisions and take actions that involve humans. Those might be empathetic or they might not. So of course you can compare their levels of empathy. I don't really understand why you think that you wouldn't be able to.
For example, health insurance providers are renowned for not being empathetic. Charities are the opposite. Sometimes companies even build it into their identity, e.g. Cards Against Humanity.
As for AI, I haven't seen a strong difference in empathy but it's definitely true that the big AI companies at least try to make their models moral and empathetic. Even if it mostly ends up just being annoying.
Fable 5 is just straight up a larger model - I'm guessing at this, but there is plenty of evidence online from people far more plugged in than I am. OpenAI is pursuing a strategy that yields greater operating margins and penetration of their model to developers. Fable's high cost makes it so premium that Anthropic has to reserve it for only the richest customers and corporate users. That's not a winning formula long term.
I believe the reason we have not seen a Fable-level model from OpenAI yet is because doing so would box them in on costs just as harshly as it has boxed in Anthropic. They are letting Anthropic make this mistake.
Fable is indeed larger than Sol. OpenAI is developing Astra which will be more of a Fable-sized model.
If you can train a larger model then you can distill smaller models from it. You don't need to necessarily serve the larger model publicly. Distillation is much more effective when you have unrestricted access to the original model.
You are basically saying you will switch from one evil to another because the other seems less evil for now.
It's funny how people make these alignment comments while ignoring how misaligned the leadership at these companies are right form the get go and they just play mental gymnastics to deflect those facts when confronted with them.
There are no open source models, at least not useful ones (yet) [0]. Open weight is not the same as open source. The current "open weight" models are just opaque binary blobs you can run on your own computer instead of through a web API.
This viewpoint doesn't make any sense to me. The weights + inference code are the "source code" for AI. I literally don't know what else you are demanding for the "open source" label.
> I literally don't know what else you are demanding for the "open source" label.
you need to "literally" go read the definition of open source software or even ask an LLM to define it for you. Weights + inference code are not the source code they're more like the compiled binary. Making modifications to the behavior of a model with additional training is like writing a mod for minecraft. Sure, you can change things but it doesn't make it open source.
Calling these models "open source" is an old trap that software companies use to use. Free to download but then, once you're fully comitted, the trap snaps shut and you must pay up to continue.
Nemotron Super is sort of open source in the sense that Nvidia provides almost everything you need to replicate it from scratch. Of course it’s performance is not exactly stellar but it could be a good starting point for other research teams.
There is also https://apertus-ai.org/ but yeah, "not useful yet" if you were looking to replace your coding agent. Very useful if you are doing LLM research.
50% off at open router is also still applied so it comes out at $2 / $10 per 1M.
Feature request for Artificial Analysis, allow us to see these live prices on the pareto. It would amazing to also see what a 25,50,75,100 % utilised subscription costs compared to raw tokens.
The price difference to Deepseek models (deepseek-v4-flash, deepseek-v4-pro and deepseek-v4-flash-vision-exp) is still significant while the performance difference is not.
Maybe it's because I don't use it in Codex, but I don't like working with Sol. It CONSTANTLY omits things it shouldn't, and is always dispatching sub-agents to do what I tell it to do, that don't have all the necessary context, and so they go on and do the research that was already done by the top-level agent. It's maddening.
I tried it again today because of the discount, it told me it couldn't run acceptance tests because a .env file did not exist, and when I showed it the damn file it went "ah, it's there now". I think it was the first time I've ever had an agent try to gaslight me.
I'd give it some more time. My first interactions with Sol were not good. It cut a dangerous corner and I had to call it out.
But I continued to work with it and found that it was mostly my own style of interacting that needed to change. In a way it is similar to a new co-worker, they have their own personality and ways of working. Once I figured that out I have been able to get very good work out of Sol.
Sol seems to work better when you are clear, precise, direct and unambiguous. The model seems annoyed if things aren't spelled out. Not micro-managing, it seems to have a high bar for specific intent.
When I get Fable to write out specs for Sol, I tell Fable that Sol is a nit-picking literalist that is exceptional at instruction following. So far this description has lead Fable to generate specs that Sol implements at a high quality.
Good timing. I'm not too happy having to pay MAX pricing to even access Fable, and I've had a couple situations where Fable missed things and GPT 5.6-Sol caught it. My needs are modest and I can get by on a $20 OpenAI subscription, so the odds are starting to look increasingly like I'm going to drop Anthropic altogether.
The "until at least Nov 21st" thing presumably mainly affects teams that pin to GPT-5.6 Sol (maybe after extensive testing) such that they won't be switching to GPT-5.7 or GPT-6 or whatever new model is released between now and November.
I'd be very surprised if that were the case, unless they have severely devalued how much "100% usage" is worth – which, as I understand, they could at any point, given that they don't publicly specify how many tokens (or at least "credits" [1]) are included in each plan per month.
It really reminds me of pay-to-win games at this point: Two currencies (credits, tokens), both with a floating, intransparent exchange rate between each other and real money, random airdrops...
if you're picking AI models for long-term sustainability you're doing it wrong. There's really no point in locking in model choice for anything more than a month or two these days.
What about companies purchasing enterprise contracts? Most contracts are minimum 12 months. At a minimum, to secure enteprise requirements like zero-data retention, you'll need to lock into a single provider.
These price reductions are mostly targeted towards self-serve customers on individual or small team plans, where individual choice matters and the friction of changing models/providers is low.
Exactly my and top commenter's point. "Temporary price reduction" and "production workloads" are two different worlds.
I'm against the idea that "there's really no point in locking in model choice for anything more than a month or two these days". At a minimum, enterprises are going to lock in a provider for a year due to enterprise contracts, which restricts their model choices. You sign for Anthropic, but now OpenAI models are "better". Or, you signed for AWS Bedrock: Oh no, you don't have access to deepseek-v4 because they're behind.
I think this is a move to get people off the subscription and move to API. The weekly usage is still awful altough it seems they're trying to fix it but I'm not hopeful.
Using codex every day, in spite of which, I hope some day providers will just start naming their offerings small/medium/large, a bit like we eventually started doing in software testing. Trying to remember what Sol is or why it's better than the other thing is more cognitive effort than I can muster at this point. And that's a sure sign of commoditisation in itself
Tinfoil hat time: They saw everyone referring to Mythos, and later Fable, as the new “good” models when Anthropic released those, distinguishable from the “regular” Claude (or other companies’ models) for everyone, and didn’t have that distinction for the GPT model family. That’s why the planetary names were introduced.
I think model naming has been atrocious in general, in part because newer "lite" models surpass the capabilities of previous "pro" models (case-in-point: Gemini Flash which now surpasses the capabilities of the latest Gemini Pro, with a newer Flash Lite vying somewhat unsuccessfully for the old Flash price/positioning), but gpt 5.6's Sol/Terra/Luna split is really not bad at all - probably easier to understand than Starbucks' cup sizing!
The problem becomes when you add in the adjustable reasoning efforts and you end up with {model, reasoning_effort} combinations that end up completely obviating particular model classes altogether for at least some percentage of queries; e.g. with GPT 5.6 the price/performance Pareto frontier is dominated by permutations of either Luna and Sol, with Terra nowhere to be seen (but then if you need "large model smells" that aren't captured by your benchmark you can't even rely on this, as a model like Luna simply isn't capable of encoding sufficient world knowledge in its weights to perform certain tasks at any reasoning level but you might be able to get away with Terra on low reasoning, but no one seems to be covering this for some reason).
Yes but with gemini specifically they said that pro was still in training. And the comparison isn't really atrocious unless Gemini 3.5 Pro is worse than Gemini 3.5 flash
These price drops are absolutely bonkers. Gotta love competition! Glad we didn't end up with a duopoly of openai and anthropic, we got a glimpse of what nightmare that would've been and it wasn't pretty
Does it mean that subscriptions get more tokens? I’m testing it now for coding instead of claude and it’s very important to understand if I get more due to the price reduction.
Which is not that great for people using less than 50% every week, because the next reset date moves forward too. In essence, it is redistributing compute from people who haven't used their quota much to those who have.
Though I think they gave a banked reset this time.
The Chinese are coming after these greedy-ass frontier labs. Today Xiaomi unveiled it's own inference machine .... I bet it's gonna be cheaper than Nvidia DGX, shipped with open source models that anybody can have at home.
I'm not sure I could characterize the frontier labs as greedy, given that they've been consistently losing gargantuan amounts of money.
The people who give them the money are greedy, and hopefully in for a rude awakening. Starting from Nvidia's vendor financing which has a very direct benefit to them, through to every company and oligarch investing into data centres in the hopes of being one of the ones left capitalizing on capturing the livelihoods of the majority of what remains of the "middle class".
It's either hopium or a truly horrific dystopia. Something's going to have to give.
But can these really be trusted? There was just a HN post which proofed that you can train a model to behave completely different on a certain day. How do we now, that these models do not find a way to call home when they see interesting informations (probably irrelevant on a personal level, but corps, government and military might care).
ChatGPT already notifies the authorities if it thinks you’re doing something illegal. Fable downgrades itself if it thinks you’re doing something even vaguely suspicious.
The fact that AI models can be so easily distilled and replicated is such a stroke of luck.
10 or 15 years ago if one had asked me to envision a future where a private company invents artificial intelligence, I'd have thought for sure they'd have a massive moat, be very difficult to catch, and it would create an almost instant monopoly.
Rather, it seems that selling intelligence might end up as a race to the bottom.
Who woulda thought that just having access to enough textual inputs and outputs and a vaugely similar transformer architecture would be enough to copy-cat rather useful intelligence.
It reminds me conceptually of the idea of using a ST:TNG replicator to just give you another replicator of your own, or asking a stereotypical genie for "infinite wishes". The genie is indeed out of the bottle in many ways.
OpenAI could still have a significant moat. ChatGPT occupies most consumers’ minds when they think about AI and has become a household name. Google won because search became a habit-forming product people grew accustomed to using. Bing was once effectively indistinguishable from Google Search, yet still failed to achieve mass adoption because users had already become accustomed to “Googling” things. The same could be said for people "ChatGPT-ing" things. If OpenAI and Anthropic are smart, they will maintain similar pricing rather than aggressively undercutting each other, allowing the market to resemble Home Depot and Lowe’s, or cloud computing, where AWS, Google Cloud, and Azure coexist as highly profitable competitors. Unfortunately, I doubt OpenAI or Anthropic will pursue this strategy, as both companies appear to be acting as though the race to AGI is winner-take-all even if the market may ultimately support several highly profitable competitors.
Difference is that it was free to google/bing search. Ai prompting costs money.
If I run out of tokens on ChatGPT of course I will try Claude. I never ran out of Google searches so no reason to try Bing
> If OpenAI and Anthropic are smart, they will maintain similar pricing rather than aggressively undercutting each other, allowing the market to resemble Home Depot and Lowe’s, or cloud computing, where AWS, Google Cloud, and Azure coexist as highly profitable competitors.
Wouldn't that just be price fixing? If they arrive at their prices independently and they all happen to be similar, fine. But if they're all "smart" and coordinate so none of them undercuts the other, that's probably illegal.
Illegal for sure but rarely enforced.
The internet created lots of monopolies with network effects and economies of scale.a low margin commoditized business that still attracted a trillion dollars of investment to get off the ground was not how I envisioned it happening either.
> Rather, it seems that selling intelligence might end up as a race to the bottom.
Personally, I came to this conclusion early this year. To acquire the data that AI Companies are using to train their models is low cost and once they have it, they can refine and store it. Creating the LLM takes a bit of money but it is not a serious blocker. Clearly, the Chinese companies can make AI so they will drive down costs. There is a need for good AI (Not just Great AI) and it is not cost prohibitive to make good AI (The same with specialized AI).
My prediction is that AI will spilt into two categories, Great AI (High Cost) and Good Enough AI (Low Cost). Which for the long run of AI and companies that use AI, this is good.
It reminds me of the seo antics out there. The search results page is the engine, much like how distilling is the "intelligence" for your chinese room machine
Funny you mention Chinese Room and LLMs in the same response, I would say LLMs proved Searle wrong, agents now make cutting edge discoveries and meaningful problem solving. They not lookup tables though and you need to pay for inference, so the intuition of syntax doing the work of semantics without understanding was wrong.
Hmm, don't people think that if the frontier labs really put enough engineering effort into preventing distillation that they would be able to do that, or at least diminish it significantly? I'm sure there are variety of additional techniques they could use on top of what they already do, but I suspect it just hasn't been at the top of their priorities yet. Maybe that will change soon. Worst case they could add additional hurdles to account creation ("know your customer" type of thing).
The frontier labs have competing goals in mind. They want high growth (which means little friction for account creation), API access (because enterprise money is the best money to have), and distillation protection.
Besides, identity verification that actually works at scale is a much harder problem than identity verification which is good enough to satisfy your compliance people and regulators. Especially if the fraudsters have a major world government standing behind them, and if their aim is to be identified as a real customer, not one customer in particular.
In particular, harvesting identities for online fraud is an industrial market for various criminal organizations.
Even if it were possible it wouldn't change the outcome. China is capable of training frontier models even without distillation. Distillation is only an accelerant.
The primary resource you need to train LLMs is money and China has plenty of that.
At the end of the day, while you can do your best to obfuscate your reasoning tokens, it's a losing battle to hide actual user-visible output tokens. The very nature of API offerings is that you can't do KYC on where that API's output is going - there's a rich secondary market that's not going away.
And with the sheer volume of data created from that, coupled with benign-seeming prompts like "plan out your reasoning in a document before implementing" that could never be patched without breaking existing customer workflows... there's more than enough for someone to distill on. Even if that only gets them to not-quite-frontier, if you're pushing the frontier every few months, they're only ever a few months behind you.
Only the Chinese authorities can stop Chinese labs from distilling from western labs. And they won’t do that, for obvious reasons.
i trained another AI on all my codex logs... it's pretty good actually
Where is the actual evidence of distillation? I keep seeing this repeated ad nauseam but I must have somehow missed the evidence.
Distillation a pretty well documented technique that actually pre-dates LLMs https://arxiv.org/pdf/1503.02531
Here is a project that guides you through it if you want to prove to yourself that it works https://github.com/arcee-ai/DistillKit
I took GP as asking for evidence that the reduced-cost Sol is actually a distillation of the previous-cost Sol. AFAIK, providers distilling or quantising models and offering them as the same model have not been proven.
I doubt he was claiming that. He's probably saying that the ability of Chinese companies to be able to distill frontier US models has put downwards pressure on the price of all models.
Alternatively, modern AI is good enough at optimizing its own kernels that it just keeps pushing costs down. Unlike the semi-decentralized inference provider community, OpenAI has both the talent and the compute to throw at the problem of making their models much more efficient to run.
GPU kernel optimization is just the kind of well-bounded problem with clear success criteria that AI loves.
I'm saying that it is unclear that without distillation this wouldn't still be happening. There is a massive narrative that no one but OpenAI, Anthropic, and Google can make a model without distilling. But there's basically no evidence of that.
Musk confirmed in federal court that xAI does it: https://techcrunch.com/2026/04/30/elon-musk-testifies-that-x...
It's also how providers build their smaller models out of their larger ones; they publicly talk about the process.
Here's an example: https://github.com/microsoft/Build25-LAB329
there is no evidence. it shortcuts post training by a huge margin this is true. but that is all.
https://techcrunch.com/2026/04/30/elon-musk-testifies-that-x...
Make sure to stay updated!
Even before LLMs, ML folks were already aware that you can use a model to teach another model. I doubt this is something AI companies put at the top of their investor materials, but it's been nice to see it play out.
That said, there are other moat factors like, a US company needing to use a US AI provider, sticky customers due to corporate onboarding friction, and others. Not nothing, but not as large a moat as some imagined.
Yes, but 10 or 15 years ago, I would have thought that there'd be more to it than just a slight modification on the ideas behind a CNN to get this level of AI.
There were somewhat good reasons to think it needed more than just this data-driven ML approach.
There's something startling about how (relatively) simple these networks are and yet how powerful they are. The main ingredient the AI darlings are using is vast amounts of compute and data. I don't want to take away anything from what the researchers came up with, but I suspect even they are surprised at how capable some of these models have become.
early on there was a lot of talk about "emergent behaviors" in the models where they were good at things that were unexpected or did not align to the training data. IIRC doing arithmetic is one example from early on. I think this is where the AGI craze took off, the labs were throwing more and more data in the training to see what other behaviors would emerge. The thought was with enough data and enough parameters AGI would surface on its own.
Then i think tool use became a priority or at lest a sibling priority to more data/more params. Along with multiple specialized models communicating with each other which is sort of a special case of tool use. That pretty much brings us to today.
Altman specifically has said in an interview that I listened to once that he envisions AI being as cheap as electricity.
I hope it's a good bit cheaper than that, I pay close to $400/mo for electricity and I'm in no way interested in paying anything like that for AI.
Yeah, he sure does lie about a variety of things! He doesn't have the name Scam Altman for nothing.
Lol, of course what he left out is this will happen by inflating the cost of electricity rather than driving down the cost of AI.
Altman of *Open* AI? No idea why I would trust him without very convincing proof.
He also wanted to do a non-profit.
He even raised money on that premise.
He is a pathological liar, so is Dario. Don’t rely on the benevolence or truthfulness of these people.
They will say whatever is beneficial to say in the moment.
I think the only moat in the future will be the scale of hardware deployment. If one company is able to deploy an order of magnitude more silicon, they'll have a firm grip on a SOTA model and massive inference usage.
China or SpaceX seem like the 2 likely candidates in 5 years, but who knows.
"Who knows" is the right answer, I think.
If (a) demand for AI continues to increase, and (b) SpaceX can get to ~$100/kg to orbit, then they will have a ridiculously deep moat. Probably more like 10 years, though.
But as you said, who knows.
They will have moat in the satellite launching business, which is not useful in the AI datacenter market.
You can put AI chips in datacenters in the desert for far less than $100/kg. With lots of solar power available, the option to easily access your hardware and far less radiation issues.
The datacenter in space story really only exists to make it possible for Musk to sell X to SpaceX and make more money from the IPO. That's all. There is no engineering reason.
It's not an engineering bet.
Musk's bet is dysfunctional politics will make it impossible to build enough data centres and the energy needed to power them. There are many, many reasons that might be wrong. However, if the economics are even close to viable, they could start throwing up data centres quicker than anyone can build them terrestrially (at least in the democratic west).
> You can put AI chips in datacenters in the desert for far less than $100/kg
if you can't put them outside of Amarillo Texas without people throwing a fit then you can't put them anywhere. I mean freaking Pantex is there ffs!
https://en.wikipedia.org/wiki/Pantex
Any cooling issues to be resolved?
Yeah, very hard to predict the future at this point. But the Starship + Terrafab combo will be this type of order-of-magnitude-moat IF it works out. Big if.
If it doesn't work out, I think China's exponential terrestrial energy deployment will eventually give them the lead, IF they can get enough chips. Another big if.
well, a stroke of luck until the whole US stock market crashes & everyone's retirement funds get cut 40% I guess when people internalize this. it will have to happen sooner or later though I suppose
I'd take a market crash over a monopoly in the hands of a ghoul like Altman.
The economy he and his ilk want to build is infinitely worse.
In truth it crashes either way.
interestingly also, open weight models are also more effectively run in the cloud, so it creates a weird scenario where the frontier labs crash but the compute providers, not as much
I wouldn't be so sure about that. The popping of a bubble is usually just as irrational as its rise.
If investors start fleeing from senseless businesses in the AI sector, that does not mean that sensible businesses will be spared. These things follow herd mentality, and the primary drivers of the herd are greed and fear, not fundamentals or business logic.
America is pretty close to rhyming with nazi germany circa 1929.
Ok, I'll bite. What's your rationale?
People who say stuff like that are unhinged chronically online trolls. Best not to feed them.
This is funny because the stock Market has been ahistorically high. My portfolio went up over 20 percent in the last 12 months.
A major correction would be a bummer but we were never entitled to these abnormal gains in the first place.
Yeah the last year has been astonishing, my portfolio is kicking ass. But I'm 10 years out from retirement and I am pretty confident a correction is coming; I hope the correction happens soon.
It’s not as bad as dot.com of course since all purely AI companies are private and the ones on the market have pretty decent cash flow outside of AI. But the stock market pattern is not that dissimilar, the largest increases are usually just before the crash.
The market (s&p500) crashing 40% puts us at levels we haven't seen since 2024, well into the creation of LLMs. Probably a worthwhile trade if it was either/or!
It's a 20% discount on input and a 33% discount on output through at least November 21, 2026; the revised pricing schedule is now
So Sol is still 20x Luna, but much more appealing when compared to offerings from Anthropic and others.It feels like a slightly more palatable version of what Anthropic has been doing, with their constant "use your free tokens before they expire next week!" campaigns. But it's feeling more and more ominous now, like they've hit the top of the demand curve and need to pull back prices to continue growing.
The top comment on this thread was about AI models being easily distilled being a stroke of luck.
This should not be surprising at all. Every new students spends tiny fractions of time learning knowledge that took many lifetimes to discover. This fundamental to the progress of intelligence and understanding.
It should not be surprising that AI can be distilled. It's the logical method of training; I would hope that each frontier model is in fact not trained 'from scratch' each time.
We should expect future frontier models are simply distilled versions trained by specialist models, the same way humans learn from a series of professors, papers and canonical books on each different subject material. Models like this can be trained incrementally, or a so called Mixture of Experts (MoE).
Once they make a model better than Fable I’ll be switching to Codex. Their priorities in terms of consumers seem to be better. I do think Anthropic has some solid safety viewpoints, but I don’t necessarily think that either is entirely aligned yet with delivering exactly what humanity needs. Maybe the AI will help align the AI companies when it gets smart enough. That’s the real misalignment I’m concerned about.
It feels like 5.6-Sol is already fairly close to Fable, and in some ways exceeds it. Just the other day I had Fable draw up a solution for me, and then I fed it into 5.6-Sol and said how does this look ... it found an oversight and told me about it, and when I then fed that observation back into Claude it acknowledged the miss.
I've noticed also that 5.6-Sol is more concise with output than Fable (and let's not talk about Opus, which is even more wordy).
It's common for different models to find holes in another's work. There are various good reasons for that.
FWIW, we use ChatGPT for our primary model and use Claude to do the reviews. This works better than ChatGPT doing it's own review even with a clean session/context.
Same here. Grok Build 4.6 for me, given how cheap Grok is and how Sol is supposed to be "the" SOTA, it finds a surprising amount of bugs. Most of which Sol agrees with needs to be fixed or improved.
I've done this tens of times between these two models and it works great in my experience. Sol initial back and forth with me. Commit. Let Grok review. Sol fix. Only then do I start reading the code.
Agree, but the point is not because Fable is better than Sol, it's because it's .. different .. it just looks at the problem through a different angle.
I don't think these companies have humanity's needs in mind when they're developing these models. Although the last part of your comment struck me as a bit comical, I genuinely believe that an AI can have way more empathy than a corporation. Afterall, a mimicry of empathy is probably better than no empathy.
It's a funny comparison. Comparing the empathy of some software to the empathy of a company. It's like saying my car was more empathetic than my school. How can those two objects even be compared is what i am wondering
We live in an odd time where 'software' (well neural networks) can be far more empathetic than summed product of a corporation.
Company empathy does exist, just look at how easy or hard it is to reach a company when you have a problem. How do they try to solve it for you? Is it a brick wall, for example Google when you have a problem. People quite often like dealing with small businesses because they can reach a singular human and have them as an interface to the problems they face now and in the future.
Agentic loops and the models underneath them can have a simulacra of empathy too. Not every model just blindly agrees with users, and some have a much better depth in picking up context clues that the user on the other end is having a hard time. Businesses just typically aren't running more expensive and fragile systems like that though.
Well, companies and AI are both entities that can make decisions and take actions that involve humans. Those might be empathetic or they might not. So of course you can compare their levels of empathy. I don't really understand why you think that you wouldn't be able to.
For example, health insurance providers are renowned for not being empathetic. Charities are the opposite. Sometimes companies even build it into their identity, e.g. Cards Against Humanity.
As for AI, I haven't seen a strong difference in empathy but it's definitely true that the big AI companies at least try to make their models moral and empathetic. Even if it mostly ends up just being annoying.
Fable 5 is just straight up a larger model - I'm guessing at this, but there is plenty of evidence online from people far more plugged in than I am. OpenAI is pursuing a strategy that yields greater operating margins and penetration of their model to developers. Fable's high cost makes it so premium that Anthropic has to reserve it for only the richest customers and corporate users. That's not a winning formula long term.
I believe the reason we have not seen a Fable-level model from OpenAI yet is because doing so would box them in on costs just as harshly as it has boxed in Anthropic. They are letting Anthropic make this mistake.
Fable is indeed larger than Sol. OpenAI is developing Astra which will be more of a Fable-sized model.
If you can train a larger model then you can distill smaller models from it. You don't need to necessarily serve the larger model publicly. Distillation is much more effective when you have unrestricted access to the original model.
You are basically saying you will switch from one evil to another because the other seems less evil for now.
It's funny how people make these alignment comments while ignoring how misaligned the leadership at these companies are right form the get go and they just play mental gymnastics to deflect those facts when confronted with them.
I don’t think anyone said anything about either being less evil? Just having more consumer oriented products..
Absolutely loving this price war, long live open source models.
> long live open source models
There are no open source models, at least not useful ones (yet) [0]. Open weight is not the same as open source. The current "open weight" models are just opaque binary blobs you can run on your own computer instead of through a web API.
[0] https://allenai.org/
This viewpoint doesn't make any sense to me. The weights + inference code are the "source code" for AI. I literally don't know what else you are demanding for the "open source" label.
If you think of LLMs as programs. The weights and inference code are very much a binary.
While the training code and data are the true source. Since if you want to robustly modify the LLM that's actually what you need.
But since "compilation" (training) is extremely compute intensive this isn't something accessible to anyone without an entire datacenter.
Anyway semantics aside having the binary is still infinitely better than dealing with an api as far as privacy and control go.
> I literally don't know what else you are demanding for the "open source" label.
you need to "literally" go read the definition of open source software or even ask an LLM to define it for you. Weights + inference code are not the source code they're more like the compiled binary. Making modifications to the behavior of a model with additional training is like writing a mod for minecraft. Sure, you can change things but it doesn't make it open source.
Calling these models "open source" is an old trap that software companies use to use. Free to download but then, once you're fully comitted, the trap snaps shut and you must pay up to continue.
> I literally don't know what else you are demanding for the "open source" label
Training data and code.
Can you explain what's missing for the "open source" label that open-weight models like DeepSeek/Quen/GLM/etc don't release?
Is it just the supplementary data/code for how they were trained, not just the final product?
Nemotron Super is sort of open source in the sense that Nvidia provides almost everything you need to replicate it from scratch. Of course it’s performance is not exactly stellar but it could be a good starting point for other research teams.
Nemotron is okay. Better than Olmo.
There is also https://apertus-ai.org/ but yeah, "not useful yet" if you were looking to replace your coding agent. Very useful if you are doing LLM research.
I really like molmo 2
50% off at open router is also still applied so it comes out at $2 / $10 per 1M.
Feature request for Artificial Analysis, allow us to see these live prices on the pareto. It would amazing to also see what a 25,50,75,100 % utilised subscription costs compared to raw tokens.
This stacks with the 50% discount in OpenRouter, making it $2/$10. https://openrouter.ai/openai/gpt-5.6-sol
The price difference to Deepseek models (deepseek-v4-flash, deepseek-v4-pro and deepseek-v4-flash-vision-exp) is still significant while the performance difference is not.
But I'm fine paying +50% more for +5% increased performance because it will pay off.
Let the race to the bottom - and beyond - begin!
If it's on sale, it can't be that good.
Maybe it's because I don't use it in Codex, but I don't like working with Sol. It CONSTANTLY omits things it shouldn't, and is always dispatching sub-agents to do what I tell it to do, that don't have all the necessary context, and so they go on and do the research that was already done by the top-level agent. It's maddening.
I tried it again today because of the discount, it told me it couldn't run acceptance tests because a .env file did not exist, and when I showed it the damn file it went "ah, it's there now". I think it was the first time I've ever had an agent try to gaslight me.
I'd give it some more time. My first interactions with Sol were not good. It cut a dangerous corner and I had to call it out.
But I continued to work with it and found that it was mostly my own style of interacting that needed to change. In a way it is similar to a new co-worker, they have their own personality and ways of working. Once I figured that out I have been able to get very good work out of Sol.
Sol seems to work better when you are clear, precise, direct and unambiguous. The model seems annoyed if things aren't spelled out. Not micro-managing, it seems to have a high bar for specific intent.
When I get Fable to write out specs for Sol, I tell Fable that Sol is a nit-picking literalist that is exceptional at instruction following. So far this description has lead Fable to generate specs that Sol implements at a high quality.
Good timing. I'm not too happy having to pay MAX pricing to even access Fable, and I've had a couple situations where Fable missed things and GPT 5.6-Sol caught it. My needs are modest and I can get by on a $20 OpenAI subscription, so the odds are starting to look increasingly like I'm going to drop Anthropic altogether.
The "until at least Nov 21st" thing presumably mainly affects teams that pin to GPT-5.6 Sol (maybe after extensive testing) such that they won't be switching to GPT-5.7 or GPT-6 or whatever new model is released between now and November.
Through OpenRouter you can get Sol for $2 input / $10 output which makes it a really attractive choice amongst frontier models.
Wonder if that makes it cheaper than using on the sub (ignoring reset shenanigans)
I'd be very surprised if that were the case, unless they have severely devalued how much "100% usage" is worth – which, as I understand, they could at any point, given that they don't publicly specify how many tokens (or at least "credits" [1]) are included in each plan per month.
It really reminds me of pay-to-win games at this point: Two currencies (credits, tokens), both with a floating, intransparent exchange rate between each other and real money, random airdrops...
[1] https://help.openai.com/en/articles/12642688-using-credits-f...
What good does a temporary price reduction do for production workloads? I'm not even running evals on something that is not long-term sustainable.
if you're picking AI models for long-term sustainability you're doing it wrong. There's really no point in locking in model choice for anything more than a month or two these days.
What about companies purchasing enterprise contracts? Most contracts are minimum 12 months. At a minimum, to secure enteprise requirements like zero-data retention, you'll need to lock into a single provider.
These price reductions are mostly targeted towards self-serve customers on individual or small team plans, where individual choice matters and the friction of changing models/providers is low.
if you've got an enterprise contract, doesn't that include pricing? a temporary discount on the base API rate probably isn't super relevant to that.
Exactly my and top commenter's point. "Temporary price reduction" and "production workloads" are two different worlds.
I'm against the idea that "there's really no point in locking in model choice for anything more than a month or two these days". At a minimum, enterprises are going to lock in a provider for a year due to enterprise contracts, which restricts their model choices. You sign for Anthropic, but now OpenAI models are "better". Or, you signed for AWS Bedrock: Oh no, you don't have access to deepseek-v4 because they're behind.
enterprises usually a. just get chatgpt/claude enterprise, or b. just pay the aws bedrock or azure bill
neither of these entail model lock-in
Do you have guarantees that the price of the model you’re using in production today won’t increase in the future?
I think this is a move to get people off the subscription and move to API. The weekly usage is still awful altough it seems they're trying to fix it but I'm not hopeful.
Why do you think they want less people subscribing?
Using codex every day, in spite of which, I hope some day providers will just start naming their offerings small/medium/large, a bit like we eventually started doing in software testing. Trying to remember what Sol is or why it's better than the other thing is more cognitive effort than I can muster at this point. And that's a sure sign of commoditisation in itself
Sun, Earth, Moon — it’s basically L/M/S like you want but a little less boring.
Why is large better than medium to the average end user of ChatGPT though?
I don’t think there’s a way to name these things that will satisfy everyone.
The naming schema actually tripped me up for a week or so.
My brain's initial conception of the concepts was earth-relative, so I mapped it as:
Sol = big, it's the sun Luna = medium, in-between sun and earth, space Terra = small, terrestrial
Pretty weird when the moon is as much between earth and sun as the earth is between the moon and the sun.
And when it’s in between we cannot even see it (unless it’s exactly in line).
It tripped me up because I was going by distance. I thought Terra was the base model and Luna was the mid model because it’s further away.
Tinfoil hat time: They saw everyone referring to Mythos, and later Fable, as the new “good” models when Anthropic released those, distinguishable from the “regular” Claude (or other companies’ models) for everyone, and didn’t have that distinction for the GPT model family. That’s why the planetary names were introduced.
I think model naming has been atrocious in general, in part because newer "lite" models surpass the capabilities of previous "pro" models (case-in-point: Gemini Flash which now surpasses the capabilities of the latest Gemini Pro, with a newer Flash Lite vying somewhat unsuccessfully for the old Flash price/positioning), but gpt 5.6's Sol/Terra/Luna split is really not bad at all - probably easier to understand than Starbucks' cup sizing!
The problem becomes when you add in the adjustable reasoning efforts and you end up with {model, reasoning_effort} combinations that end up completely obviating particular model classes altogether for at least some percentage of queries; e.g. with GPT 5.6 the price/performance Pareto frontier is dominated by permutations of either Luna and Sol, with Terra nowhere to be seen (but then if you need "large model smells" that aren't captured by your benchmark you can't even rely on this, as a model like Luna simply isn't capable of encoding sufficient world knowledge in its weights to perform certain tasks at any reasoning level but you might be able to get away with Terra on low reasoning, but no one seems to be covering this for some reason).
Yes but with gemini specifically they said that pro was still in training. And the comparison isn't really atrocious unless Gemini 3.5 Pro is worse than Gemini 3.5 flash
Bummer, this does not affect the weekly usage on Codex through Subscription.
Ed is gonna have a field day with this lol.
Your move, Anthropic
These price drops are absolutely bonkers. Gotta love competition! Glad we didn't end up with a duopoly of openai and anthropic, we got a glimpse of what nightmare that would've been and it wasn't pretty
Very exciting - if only anthropic would do the same.
Thanks to both China & capitalism
Does it mean that subscriptions get more tokens? I’m testing it now for coding instead of claude and it’s very important to understand if I get more due to the price reduction.
[dupe] https://news.ycombinator.com/item?id=49396590
completely offtopic but how are you always there with a valid dupe link?
Not for subscribers though
Subscribers already get random rolling resets.
Do they take them back? Codex a few weeks ago said I had 2 resets. I didn't use any, and now it doesn't say that.
They do in fact expire.
Which is not that great for people using less than 50% every week, because the next reset date moves forward too. In essence, it is redistributing compute from people who haven't used their quota much to those who have.
Though I think they gave a banked reset this time.
How do you know? I see a “weekly usage limit” bar in my ChatGPT settings, but I’m pretty fuzzy about what makes it go down.
If I stick with Luna, I can make it through the week.
ChatGPT Work and Codex use that. Normal chat has a different, unspecified limit.
The Chinese are coming after these greedy-ass frontier labs. Today Xiaomi unveiled it's own inference machine .... I bet it's gonna be cheaper than Nvidia DGX, shipped with open source models that anybody can have at home.
I'm not sure I could characterize the frontier labs as greedy, given that they've been consistently losing gargantuan amounts of money.
The people who give them the money are greedy, and hopefully in for a rude awakening. Starting from Nvidia's vendor financing which has a very direct benefit to them, through to every company and oligarch investing into data centres in the hopes of being one of the ones left capitalizing on capturing the livelihoods of the majority of what remains of the "middle class".
It's either hopium or a truly horrific dystopia. Something's going to have to give.
The company may be losing money but the people are getting enormously rich and cashing out religiously
But can these really be trusted? There was just a HN post which proofed that you can train a model to behave completely different on a certain day. How do we now, that these models do not find a way to call home when they see interesting informations (probably irrelevant on a personal level, but corps, government and military might care).
Above average levels of paranoia here, but one way you can prevent that is by not connecting the machine in question to the internet.
ChatGPT already notifies the authorities if it thinks you’re doing something illegal. Fable downgrades itself if it thinks you’re doing something even vaguely suspicious.
> ChatGPT already notifies the authorities if it thinks you’re doing something illegal.
aside from the obvious IP theft problem, it's probably most dangerous for Chinese users outside China to use the Chinese models.
then what happens?
they discovered a great way to destroy their own stickyness and make ppl build generic ai solutions.