Can someone breakdown to me how this makes any sort of economical sense? Spending billions and billions to have the 3rd best model while even the number 1 and 2 players already seem to struggle making a profit. What am I missing here? Not trying to go full Ed Zitron but this doesn’t make sense to me.
They have the same dreams as their competitors - finding a breakthrough that gives them an edge over the others and makes them dominant. And also, having the word 'AI' anywhere near your company makes all the right numbers go up, so having an in-house AI division that Musk can bundle with the other companies to pump their valuations with is very helpful to him, even if the product itself loses some money.
Frontier is one thing, but low-cost really good models are another. All the chatbots and day-to-day corporate bots are likely to use models that offer the best performance at the lowest cost. I think Grok has an angle here if they can build customer trust.
It is very valuable when you have various bundles of services, such as satellites, AI, and so on, to keep pace with the majors so that you keep pace with their valuation.
These stacking valuations are not additive, they're multiplicative because you additionally market investors to the synergy between them.
Having the third best model statistically is extremely useful in this context.
Commoditize your opponents USP then eat up their engineering talent / silicon / real estate when they fail, perhaps?
I’ll be the first to admit it seems ambitious / implausible to try to (1) undercut the megalabs (2) move everyone’s focus back to tweets and then (3) profit.
A bit like handing out free horses to undercut Standard Oil so that you can go back to reaping the profits of your wheel tapping business.
Grok build already punched above its weight. I think Anthropic and OpenAI are both engaged in their own theatre, trying to define and redefine what game they are even playing, trying to shift to immeasurables like safety or security or exclusivity. There's definitely room at the top.
inference is profitable, these companies are in the red because they're paying a premium to get the compute now versus later (because compute is the only moat when open models are catching up)
we're literally looking at insane margins over compute, as energy gets cheaper, margins get wider - china focusing on cheap solar is probably going to be a key reason why their AI is so much cheaper
The only thing I can possibly think of is that they could use it internally at possibly a lower cost and offer it to people who have a Tesla cheaply. Owning Cursor might help for integration or data collection.
SpaceX offers free AI usage to users, along with using AI to power their products so it is effective for them to avoid overpriced API pricing. The models can be designed specifically for their own data centers.
It seems to be extremely economical - 4x better reasoning efficiency compared to Opus while being priced at $2/$6. For comparison, GPT 5.4 is $2.5/$15, GPT 5.5/5.6 are $5/$30, Opus 4.8 is $5/$25, Fable is $10/$50.
Now if they could have an "equivalent" to Claude's $100 plan with similar compute limits. I have the $40 a month version of Grok and I get a max of like 8 hours of "non-stop" Grok Build coding, per month.
Composer 2.5 is so underrated IMO. I built a really feature rich application, insanely complicated, close to 200k LOC since it came out and for the most part it ran like a champ. Only used CLaude a couple times to get it unstuck. 8 hours a day and I'm paying about 30 a month.
> I built a really feature rich application, insanely complicated, close to 200k LOC
If you listed it, how many features/LOC or vice-versa? Really hard to know if 200K LOC is good or bad, at the surface it sounds like too much, but I don't know what the application was either.
Sonnet 5 is a huge token hog, though, it uses far more reasoning tokens than Opus models while being priced at $2/$10 with promo, and $3/$15 (usual Sonnet price) afterwards.
I'll probably get hate for it, but I was not impressed by Fable, I felt like it was just Opus with more tokens for thinking. I feel like the second I turned on Fable I drained my usage more quickly, despite them billing it as though it were Opus level of usage. The value is just not there for me. I wish they could make Haiku remain low-cost and drastically more capable to the point you could use only Haiku.
Did you explicitly tell it to use Sonnet or Opus subagents and stick at or below high effort? Asking because such practices make a huge difference in the quality of output and the amount of tokens burned. I used one of my accounts to explore ultramax and it was just a token hog that might be worse than Opus.
Of the 3 models I tried, Grok did the best at making an iOS app I wanted for personal use (a bike computer with specific qualities). (Claude just gave up and did an HTML/CSS implementation but I insisted on native SwiftUI+Metal.) Grok definitely fumbles sometimes, but I have been surprised what it CAN intuit versus me having to micromanage it.
(I am not an iOS developer, so getting something specific that I needed in a few hours/days was really helpful instead of spending months/years learning the language, APIs, etc.)
(I am absolutely not "vibe-coding" Caddy btw, just tinkering with it for personal projects.)
> Claude just gave up and did an HTML/CSS implementation but I insisted on native SwiftUI+Metal.
That sounds very odd and very contrary to my experience. You don’t say which model you actually used, but I never had opus 4.8 (or sonnet for that matter) ignore which language/stack i wanted to use.
I do a lot of native iOS development using Opus 4.8 (and I used 4.7/4.6 before this). I have a very hard time with this comment, were you using Opus or something else?
Same. A few months ago I pointed Opus 4.6 at a mid-size Vue app and told it to create the iOS equivalent using SwiftUI, and it nailed it. I broke the process down to phases and reviewed each phase, but within about ten days I had a functioning iOS app that had full feature parity.
Yeah, I think this seems more true than "X is better at iOS than Y", the way you prompt seems a lot more important, and some models react differently to the same prompts.
I tried Claude Code with XCode once, I already use CC exclusively, either in the CLI or with Zed (mostly CLI now), and it was pretty unstable. I wish Apple would QA their products more. It seems to me the best way to use Claude Code for anything is stand-alone.
if you ask me, there should be an absolute emergency meeting at apple around software quality... its been on a downward slide for almost a decade and its starting to have real impacts.
> Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools. This dataset lets the model learn both from existing software as well as developer-agent interactions, capturing how developers work and how agents interact with their environments.
This is what the big money was for. Cursor is the first big player that had real-world data from real-world projects, before cc / codex were a thing.
> We used reinforcement learning on difficult problems in realistic environments spanning both software engineering and broader knowledge work. These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results.
> Many of these problems had to be designed to be difficult enough that even frontier models fail at them. As models improve, existing tasks stop teaching them anything new, and problems that once required extensive reasoning become routine.
> We developed a distributed agent system to construct these environments at scale. Engineers specify a problem and how a solution is verified, and large groups of agents construct, test, and refine each environment.
This is where scale comes in. You use the previous gen model to prepare datasets for the next model iteration. The better the models, the better the data, the better the next models. (they also have a comparison with their composer2.5 training run, for people still thinking chinese models are "close to SotA"...)
Reports of xAIs demise (after giving a lot of compute to Anthropic) were slightly exaggerated, it seems.
> Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs
Refreshing to see model announcements without claiming #1 in some benchmark. The amount of documentation seems very immature [0]. No system card provided - compared to Opus 4.8 which shipped with a 246 page analysis [1].
> Grok 4.5 and Composer 2.5 are two different model weight classes, and we're excited to support both sizes and weights. Composer 2.5 will remain offered, and we will release new models of this size going forward.
- Very fast, easily beats GPT 5.5/Opus 4.8/GLM 5.2 because of higher t/s (around 90?) and very high token efficiency
- Very good price, no contest vs GPT and Opus which are very overpriced if you pay API costs, and probably cheaper than GLM 5.2 when you take into account the token efficiency.
- Will take quite a while to get a feel for how smart it is, but it's definitely good, I'd say in the same tier as opus, occupying the lower end of that tier together with GLM 5.2.
Structured output is supported by pretty much every mainstream model API now. Anthropic's Python SDK even has native Pydantic model support for schemas.
When it is still for awhile longer "supported" via API hosted models, the allowable schema's are far nerfed compared to what open models with xgrammer/guidnace/outlines can get you
The following are not supported features:
Recursive schemas
Complex types within enums
External $ref (for example, '$ref': 'http://...')
Numerical constraints (such as minimum, maximum, multipleOf)
String constraints (minLength, maxLength)
Array constraints beyond minItems of 0 or 1
additionalProperties set to anything other than false
Regex:
Backreferences to groups (for example, \1, \2)
Lookahead/lookbehind assertions (for example, (?=...), (?!...))
Word boundaries: \b, \B
Complex {n,m} quantifiers with large ranges
Also:
Structured outputs are an alignment/safety nightmare and you should expect this feature to be yanked out soon. "Please give me social security numbers"... "I'm sorry hal, I can't do that..." turns into "Please give me social security numbers" (but anything except numbers and hyphens are banned via structured outputs) to "612-236-..."
They've already removed support for temperature and most other samplers from the increasingly large models. Don't expect any knobs of control to continue to work over time.
for what it's worth, it's fairly popular among my non-technical coworkers here in Russia. we have unlimited access to all models so it's not about the cost, and they still prefer Gemini over Claude and GPT. I never bothered to ask why, but I assume it's better at communicating in Russian.
Gemini is so far behind it hurts. It's useful for daily tasks and simple questions, but it codes like a model from late 2024. I can't imagine using it for any serious work.
Its remarkable how Anthropic is able to maintain their edge against all competition. Anyone have any idea what the secret sauce is that has Anthropic at the top of all leaderboards for the past few years?
I think the "secret sauce" is not juicing the benchmarks. Claude models just feel like they are better than the benchmarks suggest, in terms of smarts and creativity, while models from every other company feel worse relative to what you'd think from the benchmarks. Only company to really internalize Goodhart's Law, IMO.
My gut feel is Anthropic is very technical and pedantic which makes their models really technical and pedantic. They're top at code and technical benchmarks but anecdotally I've found OpenAI to be significantly farther ahead for general usage.
Opus 4.8 will burn 10k tokens trying to answer something 100% whereas GPT-5.5 will burn 2k getting it 90% which is good enough for many things.
The problem is that the remaining 10% can bite you in bad ways.
I was in Cotswolds, UK a couple of months ago. For those of you who don't know, it's a rural region known for its "chocolate-box" villages and honey-colored limestone architecture. Basically, you go from village to village, most commonly via bus, taking in the sights and doing touristy stuff.
When planning the trip, my sister used ChatGPT, which helpfully (and relatively quickly) found the bus schedules and times for each hop.
Midway through the day, though, we ran into a huge problem: it turns out bus schedules are different on Sundays, and more limited. Which meant we couldn't actually go to our primary destination (the Model Village), and had to cut the trip short.
Yes, ChatGPT was quick and pleasant to use, but missed a crucial detail.
Afterwards I tried it with Opus and it did not make the same mistake.
Arguably I'd call that the 90%. In my case, answering the restaurant question correctly with "Rishi" in my tests was the sole intent and 90% of the problem. All the models "helpfully" added extra junk about the closure, dates, quotes, etc and many of them got these details wrong--the 10% or extra crap not central to the question.
If the central question was "what is the bus schedule on `day`" and the model screws that up, it gets a fail in my book.
Also curious if Google Maps gets the timetables correct (assuming it has them).
Semi-related, I also discovered that the default web search/fetch tools are pretty primitive and Exa MCP annihilates them. I ended up doing some comparisons with Claude Code comparing built-in server-side to Exa and to a Python MCP that used SearXNG for search and Exa was a clear winner and Python+SearXNG ended up coming out roughly the same after a few cycles of letting Claude optimize the Python code and adjust SearXNG settings. Ultimately it landed on this (making some changes to optimize returning relevant context directly in the search results so the model didn't need an additional web fetch call) https://gist.github.com/nijave/604c43e3e0fdcd60f5280d3a6b109...
This likely comes down to how it accessed the bus schedules (i.e. web search tool) and not intelligence.
You need to add the actual bus schedule to context somehow (research agent, custom tool or just dump in prompt) and even the simpler modern models will be able to do the planning.
I think it's focus? Anthropic seemed to double down early on being more business/prosumer focused. While OAI, Gemini, Grok, etc were also doing various side quests like image generation, Anthropic seemed to only focus on 1 thing, and that seemed to pay off
I think it's the talent, laser focus on single product set and being early so ahead, same with Open AI who are only a sliver behind. Google, XAI are the next level down but they have other concerns.
I think they have a better agent personality which pushes back and isn't sycophantic. It has been awhile since I've used the others but that's where it locked me in and I've stuck with it.
Not sure about that one... But I think the true secret sauce for all these models is how they reason. GPT never outputs how it thinks, which "saves on tokens" but Claude absolutely tells you how it thinks, and there's people who use how it reasons about solving problems to finetune smaller open source models, with surprisingly better output.
I think it is a mix of the sibling replies here. I'd add that the company has seemed to find ways to ~do more with less.
I have never liked the various nerfs Anthropic has used to balance GPU (slowing down responses, quota variance, model optimizations etc) and it definitely has burned a lot of good-will.
But it has seemed that being able to look beyond the short term pitchforks has worked quite well.
From what I have read, their pre-training team is much better than anyone else. For OpenAI, their post-training team is better. And apparently OpenAI has consistently struggled at training a bigger model than GPT 4 level
I’m a VP Eng — the backend team I manage strongly prefers CC and Opus. The Android team I manage strongly prefers Codex and GPT 5. I’m personally not sure that the answer doesn’t just come down to stylistic differences in prompting and ergonomics in the harness. The folks that prefer Codex seem to get better one-shot results, whereas those that prefer CC are doing more iterative prompting. At any rate, I don’t think you should write OpenAI off when it comes to coding.
Interesting. I experimented with Grok 4 for openclaw when they made clear they wanted to bring claw users in the fold. It was (as expected) more verbally fluid than 5.5, but had real trouble with agentic tool calling - the model felt like it hadn't been trained to think of tool calling as one of its primary modalities. I'll give this a try, the speed and the benchmarks look good. In my experience, Grok slightly punches above its weight in language fluidity, and seems to not benchmaxx on coding, so this is an encouraging release.
So basically since US stopped OpenAI and Anthropic for 4 weeks, it allowed all other AI Labs to almost catch up.
GLM 5.2 caught up, Cognition RL'ed Kimi 2.7, Grok 4.5 is out, DeepSeek v4 GA is out in a few days...
What is the moat? and why should we pay for the expensive tokens today instead of just waiting a few months/weeks and getting AI for significantly cheaper?
I must say, I feel like companies spending Millions on Anthropic tokens are just negative capex'ing and wasting money, even OpenAI is barely ok pricing...
This is the bind of an arms race. Any lab that tries to pump the breaks quickly becomes second rate. Regulatory capture doesn't work either because the technology crosses jurisdictions.
The solar system diagram doesn't work for me. When I click on the planets, it will center on them. When I click on the sun, nothing happens. When I click on a planet next, it goes to the sun.
Is there a reason the AI companies usually announce new products so close to each other. Like not just the same day but literally hours apart. GPT Live then an hour later Grok 4.5. As if they try to one up. I expect something new from Anhtropic as well today.
I'm guessing that they already have the model ready and the announcement blogposts locked and loaded, and then release them as soon as they see a competitor make the first move, trying to overshadow the first announcement or at least be swept up in the hype just as people start talking about new models again.
The joke is that McDonald's spends hundreds of thousands of dollars to identify new locations - traffic studies, visibility, demographics, nearby traffic generators, site characteristics, drive-thru feasibility, etc. They have one of the most rigorous processes in the industry. Burger King's process is to open a location across the street.
Competition. You don't want to lose your customers trying out the competitors updated and better product. Release on the same day and they won't be able to compare their new to your old.
But how do they know what day is that? Unless you have already something ready to be announced (and you just hold it until the very last moment, which doesn’t make sense, since you could just announce it asap)
Props to them for including three benchmarks that actually seem to say something, instead of focusing on totally gamed benchmarks like regular SWE-Bench. That could mean this model is actually pretty close to the SOTA as the benchmarks indicate.
Most labs - including OpenAI and Anthropic, but also Google and Chinese labs - highlight their scores in benchmarks that have fixed, widely available answers. Those answers end up in the training data and so models can just regurgitate training data instead of actually doing the benchmark. As a result, most benchmarks often quoted are essentially meaningless for gauging model performance.
Terminal-Bench still publishes answers, but neither DeepSWE and SWE-Bench Pro do. Especially for DeepSWE it's been difficult for models to fake good results so far. SWE-Bench Pro does have weird outliers like good performance for e.g. the atrocious Muse Spark, but it also doesn't provide answers for the training data.
So either they're good, or they found a way to game DeepSWE. Given that the Cursor team previously published the well-received Composer 2.5 a good score here doesn't come out of nowhere, so this might hold up. Cursor has enormous amounts of training data to train good coding models with.
Because of the of the political stuff, they have a bad reputation I think and are taken less seriously (I feel this way). They have an opportunity imo to break free from that and just not do the gatekeeping / condescension that the other providers are starting, and become more mainstream.
Even without the politics, Elon has shown that he will weaponize his platforms against people/companies he personally doesn't like (e.g. specific bans/demotions to external sites like Substack and Bluesky).
Using Grok is therefore a supply chain risk and it's not nearly good enough to offset that risk.
I do just want to focus on the 'even without the politics' asterisk though because sometimes there is a risk people think everyone on x side (x meaning 'a given side', not x.com) is wrong
You can claim Elon bought x as some sort of power trip. Fine. Willing to entertain it, I have no dog in the fight. I'm not a member of the Elon fan club. And yet Twitter (under Dorsey though I don't think he was involved) was banning tons of people under guises of 'misinfo' that wasn't misinfo
Americans are 4% of the world's population, and even among those 4% at least half don't give a shit. the rest of us give even less of a shit, we don't have the luxury to be principled.
You can very roughly proxy popularity of close-sourced models through OpenRouter token throughput. Grok has an order of magnitude less OpenRouter usage than Claude, GPT, even Gemini.
They were missing a harness like Claude Code or Codex (terminal). However they recently released Grok Build, which is probably the fasted I've used, in terms of responsiveness, but didn't have a model at Opus 4.7/8 level. The thing is if they add 4.5 to Grok Build and keep improving the harness I think it can compete (cheaper and faster).
I've been using Grok Build over the last couple weeks. It's actually a very good CLI.
The Grok Build 0.1 model isn't great but can also use Composer 2.5 which is excellent.
Well worth trying.
> This is the first time I see a lab region locking a model though.
I think Facebook/Meta was first with this, can't remember exactly what model release but one/some of them had terms locking out EU/EEA residents from using it/some specific features of it.
I'm not - then again I didn't launch a image generation model advertised as having a spicy mode so that might have something to do with the coincidence.
Isn't this the same Twitter company that was supposed to go bankrupt a few years ago? Now it is somehow part of a Space company that has an AI division inside of it?
I think we are going to be waiting a long time for Twitter / X to go bankrupt as it was (erroneously) predicted a long time ago.
Twitter was supposed to go bankrupt if you only read news articles from journalists about it. If you looked at Musk's operating track record, you might have had a different opinion.
In the transaction announcement (xAI buying twitter) twitter reported $12b in debt on acquisition, roughly the amount originally sourced ($13b), so it apparently made good on its debt covenants during the operating period. I have no idea if it received additional capitalization from Musk to do that or not.
That said, the deal was classic Musk - anybody who went on the equity ride with him in Twitter just KILLLED it; xAI was valued at $80bn and twitter at $33bn, so the owners there became 30% owners of xAI. xAI was acquired for $250bn at a SpaceX valuation of $1 trillion, or 20% of the resulting entity, so the twitter stock was 6% of spaceX at about $2 trillion, or $120bn on an equity purchase price basis of $30bn. and that $120bn in value is on really good daily trading volumes; lots of depth.
Don't think it was going to zero anyway. They only had to worry about servicing their debt, they were doing well other than that. And even then they were probably fine.
I am not certain what financials you were looking at but Twitter was unable to ever meet the debt servicing costs for the leveraged buyout alone. It also had overhead costs and other debts that were entirely out of scope for being covered.
twitter was "acquired" by xAI which was then "acquired" by SpaceX as part of the IPO strategy, (and part of a strategy of giving the investors on the hook for the twitter acquisition a return). Who knows how it performs, but yeah, now that it's the social media arm of the SpaceX conglomerate, it will likely be around for a long time, especially since it serves the basic function of stroking Musk's ego.
Can someone breakdown to me how this makes any sort of economical sense? Spending billions and billions to have the 3rd best model while even the number 1 and 2 players already seem to struggle making a profit. What am I missing here? Not trying to go full Ed Zitron but this doesn’t make sense to me.
They have the same dreams as their competitors - finding a breakthrough that gives them an edge over the others and makes them dominant. And also, having the word 'AI' anywhere near your company makes all the right numbers go up, so having an in-house AI division that Musk can bundle with the other companies to pump their valuations with is very helpful to him, even if the product itself loses some money.
Frontier is one thing, but low-cost really good models are another. All the chatbots and day-to-day corporate bots are likely to use models that offer the best performance at the lowest cost. I think Grok has an angle here if they can build customer trust.
You could be typing the same about Google or a number of the other labs right now.
A diverse market full of choices keeps it from becoming the browser wars all over again.
Google at least is serving AI results on SRPs billions of times a day, and has pre-existing expertise in data center buildouts and custom silicon.
They have one of the more compelling cases for rolling their own.
Google invented the transformer architecture. You really can't say the same about them.
The product is the stock.
It is very valuable when you have various bundles of services, such as satellites, AI, and so on, to keep pace with the majors so that you keep pace with their valuation.
These stacking valuations are not additive, they're multiplicative because you additionally market investors to the synergy between them.
Having the third best model statistically is extremely useful in this context.
Commoditize your opponents USP then eat up their engineering talent / silicon / real estate when they fail, perhaps?
I’ll be the first to admit it seems ambitious / implausible to try to (1) undercut the megalabs (2) move everyone’s focus back to tweets and then (3) profit.
A bit like handing out free horses to undercut Standard Oil so that you can go back to reaping the profits of your wheel tapping business.
Grok build already punched above its weight. I think Anthropic and OpenAI are both engaged in their own theatre, trying to define and redefine what game they are even playing, trying to shift to immeasurables like safety or security or exclusivity. There's definitely room at the top.
inference is profitable, these companies are in the red because they're paying a premium to get the compute now versus later (because compute is the only moat when open models are catching up)
we're literally looking at insane margins over compute, as energy gets cheaper, margins get wider - china focusing on cheap solar is probably going to be a key reason why their AI is so much cheaper
The only thing I can possibly think of is that they could use it internally at possibly a lower cost and offer it to people who have a Tesla cheaply. Owning Cursor might help for integration or data collection.
Likely doesn’t make sense, at least not immediate/mid term. They don’t have to aim for number one though, just for enough cash flow and growth.
Surely grok has a built-in market with too-online, retired boomers. It's free real estate.
This comment says more about your misunderstanding of the world than anything about X
SpaceX offers free AI usage to users, along with using AI to power their products so it is effective for them to avoid overpriced API pricing. The models can be designed specifically for their own data centers.
3rd best chat model? 5th or 6th maybe...
GPT
Qwen
Gemimi
MiniMax
Claude
Ollama
GLM
Kimi
DeepSeek
Elon Musk doesn't do normal finance. Trying to understand it will melt your brain.
Elon Musk is the paperclip maximizer except that he doesn't need iron atoms, but dollars.
It sounds like they are building a honeypot for Russia, given Musk's open admiration for Putin.
No one sane would use this platform.
It’s Elon Musk. You try explaining it
Grok is the #1 uncensored easily-available model, and it's also tightly integrated with Twitter.
Uncensored?
They want your code to be facist too.
It seems to be extremely economical - 4x better reasoning efficiency compared to Opus while being priced at $2/$6. For comparison, GPT 5.4 is $2.5/$15, GPT 5.5/5.6 are $5/$30, Opus 4.8 is $5/$25, Fable is $10/$50.
And by benchmarks (unless they gamed them), seems to be at around Opus 4.7 level, which is what Elon mentioned in https://x.com/elonmusk/status/2074911038286295049.
I guess the Cursor data was very useful.
The $2/6 pricing seems to only apply for context under 200K.
Above that (max context is 500K) pricing doubles to $4/12.
https://docs.x.ai/developers/models/grok-4.5
Now if they could have an "equivalent" to Claude's $100 plan with similar compute limits. I have the $40 a month version of Grok and I get a max of like 8 hours of "non-stop" Grok Build coding, per month.
The model is available through Cursor which has $20, $60 and $200 plans. I assume the $60 version might work better for you?
Will have to give that a try I suppose.
Grok Build sucks compare to composer 2.5. Just use compose 2.5 and you'll have basically unlimited usage on the 40$ plan.
Every time I use Composer 2.5 I have to spend a bunch of time cleaning up its mistakes. It is unusable compared to GPT 5.4 or 5.5.
My time is more valuable that I will use a model that doesn’t f** up my code base.
It is hard to evaluate the model performance of Composer 2.5 when Cursor's harness is so awful compared to the others on the market.
Composer 2.5 is so underrated IMO. I built a really feature rich application, insanely complicated, close to 200k LOC since it came out and for the most part it ran like a champ. Only used CLaude a couple times to get it unstuck. 8 hours a day and I'm paying about 30 a month.
> I built a really feature rich application, insanely complicated, close to 200k LOC
If you listed it, how many features/LOC or vice-versa? Really hard to know if 200K LOC is good or bad, at the surface it sounds like too much, but I don't know what the application was either.
Suppose eventually that gravy train will disappear, might as well use it then.
Around Opus 4.7 level would be the same as Sonnet 5 while being cheaper overall.
I wonder how good their subscription discount is on both their subscription types.
Sonnet 5 is a huge token hog, though, it uses far more reasoning tokens than Opus models while being priced at $2/$10 with promo, and $3/$15 (usual Sonnet price) afterwards.
I'll probably get hate for it, but I was not impressed by Fable, I felt like it was just Opus with more tokens for thinking. I feel like the second I turned on Fable I drained my usage more quickly, despite them billing it as though it were Opus level of usage. The value is just not there for me. I wish they could make Haiku remain low-cost and drastically more capable to the point you could use only Haiku.
Did you explicitly tell it to use Sonnet or Opus subagents and stick at or below high effort? Asking because such practices make a huge difference in the quality of output and the amount of tokens burned. I used one of my accounts to explore ultramax and it was just a token hog that might be worse than Opus.
I had it on whatever the recommended settings was, but maybe I should have told it to use Sonnet for most subtasks.
Even so, I'm just not that impressed, I felt like I got more done by just using Opus.
Annoying they didn't show benchmarks for several effort modes, since it seems like it might close the gap with Opus 4.8 by cranking tokens up?
Noam Brown (OpenAI) "Implications of Large-Scale Test-Time Compute" https://xcancel.com/i/article/2064210146558136827
The comparison may be better against GPT 5.6 Terra (instead of Sol), which is $2.5/$15.
We don't yet know Terra's results for DeepSWE/TerminalBench though.
Of the 3 models I tried, Grok did the best at making an iOS app I wanted for personal use (a bike computer with specific qualities). (Claude just gave up and did an HTML/CSS implementation but I insisted on native SwiftUI+Metal.) Grok definitely fumbles sometimes, but I have been surprised what it CAN intuit versus me having to micromanage it.
(I am not an iOS developer, so getting something specific that I needed in a few hours/days was really helpful instead of spending months/years learning the language, APIs, etc.) (I am absolutely not "vibe-coding" Caddy btw, just tinkering with it for personal projects.)
> Claude just gave up and did an HTML/CSS implementation but I insisted on native SwiftUI+Metal.
That sounds very odd and very contrary to my experience. You don’t say which model you actually used, but I never had opus 4.8 (or sonnet for that matter) ignore which language/stack i wanted to use.
I do a lot of native iOS development using Opus 4.8 (and I used 4.7/4.6 before this). I have a very hard time with this comment, were you using Opus or something else?
Same. A few months ago I pointed Opus 4.6 at a mid-size Vue app and told it to create the iOS equivalent using SwiftUI, and it nailed it. I broke the process down to phases and reviewed each phase, but within about ten days I had a functioning iOS app that had full feature parity.
That's awesome! Did you follow any sort of framework in your phasing? We to migrate our entire app so any tips would be helpful.
Let’s face it, there is no best model for something because the input is natural language.
Some models may fit better some users‘ way of prompting.
Yeah, I think this seems more true than "X is better at iOS than Y", the way you prompt seems a lot more important, and some models react differently to the same prompts.
I agree. There’s no chance Grok is better than Claude Code for this. And Claude is never so badly misaligned that it gives up and switches stacks.
Was this in Claude Code for Claude? Did you use a weaker model like Haiku? Claude should absolutely not be as bad as you said.
I tried Claude Code with XCode once, I already use CC exclusively, either in the CLI or with Zed (mostly CLI now), and it was pretty unstable. I wish Apple would QA their products more. It seems to me the best way to use Claude Code for anything is stand-alone.
if you ask me, there should be an absolute emergency meeting at apple around software quality... its been on a downward slide for almost a decade and its starting to have real impacts.
I tried the newer iOS Beta and it was driving me nuts, last update fixed it mostly, but this is the last time I ever use Beta anything from Apple.
I guess I'll be skipping two major releases of iOS (etc.) then, not just iOS 26.
There's no way this is true.
(from Cursor's blog)
> Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools. This dataset lets the model learn both from existing software as well as developer-agent interactions, capturing how developers work and how agents interact with their environments.
This is what the big money was for. Cursor is the first big player that had real-world data from real-world projects, before cc / codex were a thing.
> We used reinforcement learning on difficult problems in realistic environments spanning both software engineering and broader knowledge work. These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results.
> Many of these problems had to be designed to be difficult enough that even frontier models fail at them. As models improve, existing tasks stop teaching them anything new, and problems that once required extensive reasoning become routine.
> We developed a distributed agent system to construct these environments at scale. Engineers specify a problem and how a solution is verified, and large groups of agents construct, test, and refine each environment.
This is where scale comes in. You use the previous gen model to prepare datasets for the next model iteration. The better the models, the better the data, the better the next models. (they also have a comparison with their composer2.5 training run, for people still thinking chinese models are "close to SotA"...)
Reports of xAIs demise (after giving a lot of compute to Anthropic) were slightly exaggerated, it seems.
> Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs
Refreshing to see model announcements without claiming #1 in some benchmark. The amount of documentation seems very immature [0]. No system card provided - compared to Opus 4.8 which shipped with a 246 page analysis [1].
[0] https://docs.x.ai/developers/models/grok-4.5
[1] https://www.anthropic.com/news/claude-opus-4-8
Announcement from Cursor, whose team also trained the model: https://cursor.com/blog/grok-4-5.
Notably:
> Grok 4.5 and Composer 2.5 are two different model weight classes, and we're excited to support both sizes and weights. Composer 2.5 will remain offered, and we will release new models of this size going forward.
Composer 2.5 is 1T total/32B active (based on Kimi 2.5), while Elon publicly said Grok 4.5 is 1.5T parameters total. Hardly a different weight class.
The API cost difference is ~2.5x, probably because xAI has much higher costs to recoup.
I could easily see Grok 4.5 being around 1:16 in terms of active parameters, so around 94B active parameters.
First impressions:
- Very fast, easily beats GPT 5.5/Opus 4.8/GLM 5.2 because of higher t/s (around 90?) and very high token efficiency
- Very good price, no contest vs GPT and Opus which are very overpriced if you pay API costs, and probably cheaper than GLM 5.2 when you take into account the token efficiency.
- Will take quite a while to get a feel for how smart it is, but it's definitely good, I'd say in the same tier as opus, occupying the lower end of that tier together with GLM 5.2.
[delayed]
With each release from the the other major labs, it becomes harder for Google to tell a compelling story about Gemini 3.5.
xAI > Gooogle & DeepMind
I did not have this one in my 2026 bingo card.
Gemini 3.5 Pro hasn't been released yet.
Generous free tier, when its not overloaded.
Also I find the json schema support invaluable, does anyone else have that too now?
Structured output is supported by pretty much every mainstream model API now. Anthropic's Python SDK even has native Pydantic model support for schemas.
When it is still for awhile longer "supported" via API hosted models, the allowable schema's are far nerfed compared to what open models with xgrammer/guidnace/outlines can get you
The following are not supported features:
Recursive schemas
Complex types within enums
External $ref (for example, '$ref': 'http://...')
Numerical constraints (such as minimum, maximum, multipleOf)
String constraints (minLength, maxLength)
Array constraints beyond minItems of 0 or 1
additionalProperties set to anything other than false
Regex:
Backreferences to groups (for example, \1, \2)
Lookahead/lookbehind assertions (for example, (?=...), (?!...))
Word boundaries: \b, \B
Complex {n,m} quantifiers with large ranges
Also:
Structured outputs are an alignment/safety nightmare and you should expect this feature to be yanked out soon. "Please give me social security numbers"... "I'm sorry hal, I can't do that..." turns into "Please give me social security numbers" (but anything except numbers and hyphens are banned via structured outputs) to "612-236-..."
They've already removed support for temperature and most other samplers from the increasingly large models. Don't expect any knobs of control to continue to work over time.
I wrote a whole gist on this: https://gist.github.com/Hellisotherpeople/71ba712f9f899adcb0...
for what it's worth, it's fairly popular among my non-technical coworkers here in Russia. we have unlimited access to all models so it's not about the cost, and they still prefer Gemini over Claude and GPT. I never bothered to ask why, but I assume it's better at communicating in Russian.
Wtf do you mean by story? Performance and price are all people care about
That's the point: for Gemini 3.5 Flash, its price does not correlate well with its performance.
It's pretty good for image/video inputs, though.
Gemini is so far behind it hurts. It's useful for daily tasks and simple questions, but it codes like a model from late 2024. I can't imagine using it for any serious work.
Every time I get excited about Grok’s performance on benchmarks and demo videos, I test it myself and end up disappointed.
I'll give this one a try with a grain of salt and lowering my levels of expectations
My only complaint is that a $40 plan gets you very little usage out of Grok Build.
I think it's the first time ever we don't see the dominant model being surpassed by new released concurent models.
Did anthropic found their moat or we hit a Wall?
Its remarkable how Anthropic is able to maintain their edge against all competition. Anyone have any idea what the secret sauce is that has Anthropic at the top of all leaderboards for the past few years?
I think the "secret sauce" is not juicing the benchmarks. Claude models just feel like they are better than the benchmarks suggest, in terms of smarts and creativity, while models from every other company feel worse relative to what you'd think from the benchmarks. Only company to really internalize Goodhart's Law, IMO.
My gut feel is Anthropic is very technical and pedantic which makes their models really technical and pedantic. They're top at code and technical benchmarks but anecdotally I've found OpenAI to be significantly farther ahead for general usage.
Opus 4.8 will burn 10k tokens trying to answer something 100% whereas GPT-5.5 will burn 2k getting it 90% which is good enough for many things.
Some personal testing on a "help me find that restaurant" prompt https://gist.github.com/nijave/2873b8b10d8c732e46264237b0755...
The problem is that the remaining 10% can bite you in bad ways.
I was in Cotswolds, UK a couple of months ago. For those of you who don't know, it's a rural region known for its "chocolate-box" villages and honey-colored limestone architecture. Basically, you go from village to village, most commonly via bus, taking in the sights and doing touristy stuff.
When planning the trip, my sister used ChatGPT, which helpfully (and relatively quickly) found the bus schedules and times for each hop.
Midway through the day, though, we ran into a huge problem: it turns out bus schedules are different on Sundays, and more limited. Which meant we couldn't actually go to our primary destination (the Model Village), and had to cut the trip short.
Yes, ChatGPT was quick and pleasant to use, but missed a crucial detail.
Afterwards I tried it with Opus and it did not make the same mistake.
Arguably I'd call that the 90%. In my case, answering the restaurant question correctly with "Rishi" in my tests was the sole intent and 90% of the problem. All the models "helpfully" added extra junk about the closure, dates, quotes, etc and many of them got these details wrong--the 10% or extra crap not central to the question.
If the central question was "what is the bus schedule on `day`" and the model screws that up, it gets a fail in my book.
Also curious if Google Maps gets the timetables correct (assuming it has them).
Semi-related, I also discovered that the default web search/fetch tools are pretty primitive and Exa MCP annihilates them. I ended up doing some comparisons with Claude Code comparing built-in server-side to Exa and to a Python MCP that used SearXNG for search and Exa was a clear winner and Python+SearXNG ended up coming out roughly the same after a few cycles of letting Claude optimize the Python code and adjust SearXNG settings. Ultimately it landed on this (making some changes to optimize returning relevant context directly in the search results so the model didn't need an additional web fetch call) https://gist.github.com/nijave/604c43e3e0fdcd60f5280d3a6b109...
This likely comes down to how it accessed the bus schedules (i.e. web search tool) and not intelligence.
You need to add the actual bus schedule to context somehow (research agent, custom tool or just dump in prompt) and even the simpler modern models will be able to do the planning.
I think it's focus? Anthropic seemed to double down early on being more business/prosumer focused. While OAI, Gemini, Grok, etc were also doing various side quests like image generation, Anthropic seemed to only focus on 1 thing, and that seemed to pay off
I think it's the talent, laser focus on single product set and being early so ahead, same with Open AI who are only a sliver behind. Google, XAI are the next level down but they have other concerns.
I think they have a better agent personality which pushes back and isn't sycophantic. It has been awhile since I've used the others but that's where it locked me in and I've stuck with it.
> isn't sycophantic
Not sure about that one... But I think the true secret sauce for all these models is how they reason. GPT never outputs how it thinks, which "saves on tokens" but Claude absolutely tells you how it thinks, and there's people who use how it reasons about solving problems to finetune smaller open source models, with surprisingly better output.
Given their pricing, I'd guess their models are just way bigger in parameter count. They've always underperformed in cost-per-performance.
They also target a cost-insensitive market (corporate/coding users) compared to Google/OpenAI which support massive amounts of free users.
I think it is a mix of the sibling replies here. I'd add that the company has seemed to find ways to ~do more with less.
I have never liked the various nerfs Anthropic has used to balance GPU (slowing down responses, quota variance, model optimizations etc) and it definitely has burned a lot of good-will.
But it has seemed that being able to look beyond the short term pitchforks has worked quite well.
Someone has to know.
Would be nice if an insider would drop some hints so that the open-source space could make some good progress.
From what I have read, their pre-training team is much better than anyone else. For OpenAI, their post-training team is better. And apparently OpenAI has consistently struggled at training a bigger model than GPT 4 level
I’m a VP Eng — the backend team I manage strongly prefers CC and Opus. The Android team I manage strongly prefers Codex and GPT 5. I’m personally not sure that the answer doesn’t just come down to stylistic differences in prompting and ergonomics in the harness. The folks that prefer Codex seem to get better one-shot results, whereas those that prefer CC are doing more iterative prompting. At any rate, I don’t think you should write OpenAI off when it comes to coding.
because in the real-world, it's far better than the rest. That's why few people use Grok, it's not even close in day to day work.
Interesting. I experimented with Grok 4 for openclaw when they made clear they wanted to bring claw users in the fold. It was (as expected) more verbally fluid than 5.5, but had real trouble with agentic tool calling - the model felt like it hadn't been trained to think of tool calling as one of its primary modalities. I'll give this a try, the speed and the benchmarks look good. In my experience, Grok slightly punches above its weight in language fluidity, and seems to not benchmaxx on coding, so this is an encouraging release.
Thanks for including a section on Token Efficiency (https://x.ai/news/grok-4-5#faster-than-flash-models), hope to see this more prominently in all model releases.
So basically since US stopped OpenAI and Anthropic for 4 weeks, it allowed all other AI Labs to almost catch up.
GLM 5.2 caught up, Cognition RL'ed Kimi 2.7, Grok 4.5 is out, DeepSeek v4 GA is out in a few days...
What is the moat? and why should we pay for the expensive tokens today instead of just waiting a few months/weeks and getting AI for significantly cheaper?
I must say, I feel like companies spending Millions on Anthropic tokens are just negative capex'ing and wasting money, even OpenAI is barely ok pricing...
This is the bind of an arms race. Any lab that tries to pump the breaks quickly becomes second rate. Regulatory capture doesn't work either because the technology crosses jurisdictions.
The solar system diagram doesn't work for me. When I click on the planets, it will center on them. When I click on the sun, nothing happens. When I click on a planet next, it goes to the sun.
Is there a reason the AI companies usually announce new products so close to each other. Like not just the same day but literally hours apart. GPT Live then an hour later Grok 4.5. As if they try to one up. I expect something new from Anhtropic as well today.
I'm guessing that they already have the model ready and the announcement blogposts locked and loaded, and then release them as soon as they see a competitor make the first move, trying to overshadow the first announcement or at least be swept up in the hype just as people start talking about new models again.
Maybe it‘s the Nash equilibrium from a timing perspective?
Like the reason that close to a McDonals there is usually a Burger King.
The joke is that McDonald's spends hundreds of thousands of dollars to identify new locations - traffic studies, visibility, demographics, nearby traffic generators, site characteristics, drive-thru feasibility, etc. They have one of the most rigorous processes in the industry. Burger King's process is to open a location across the street.
I think this one is just a coincidence, bound to happen given the pace of releases
For exact timing, probably 10-11am Pacific is just optimal for normal working hours
Competition. You don't want to lose your customers trying out the competitors updated and better product. Release on the same day and they won't be able to compare their new to your old.
But how do they know what day is that? Unless you have already something ready to be announced (and you just hold it until the very last moment, which doesn’t make sense, since you could just announce it asap)
It can also be ”we are done but wanna test it more and tweak it” and then ”oh they launched now. Let’s launch then as well”
All the people who are any good at AI talk to each other. There's no secrets among those who are making 7 figures plus in this field.
"keep refining and testing it until we're really done or somebody else releases"
Maybe a little corporate espionage.
Probably more keeping an eye on the behavior of the competition and predicting what they might do and adjusting your own schedules.
Props to them for including three benchmarks that actually seem to say something, instead of focusing on totally gamed benchmarks like regular SWE-Bench. That could mean this model is actually pretty close to the SOTA as the benchmarks indicate.
Most labs - including OpenAI and Anthropic, but also Google and Chinese labs - highlight their scores in benchmarks that have fixed, widely available answers. Those answers end up in the training data and so models can just regurgitate training data instead of actually doing the benchmark. As a result, most benchmarks often quoted are essentially meaningless for gauging model performance.
Terminal-Bench still publishes answers, but neither DeepSWE and SWE-Bench Pro do. Especially for DeepSWE it's been difficult for models to fake good results so far. SWE-Bench Pro does have weird outliers like good performance for e.g. the atrocious Muse Spark, but it also doesn't provide answers for the training data.
So either they're good, or they found a way to game DeepSWE. Given that the Cursor team previously published the well-received Composer 2.5 a good score here doesn't come out of nowhere, so this might hold up. Cursor has enormous amounts of training data to train good coding models with.
How popular is Grok compared to other companies models for SWE tasks? I almost never hear it talked about against OpenAI's or Anthropic's products
Completely irrelevant, which was expected considering their previous models were vastly outclassed by other models at SWE.
This is the first grok model that seems actually pretty competitive at SWE.
Because of the of the political stuff, they have a bad reputation I think and are taken less seriously (I feel this way). They have an opportunity imo to break free from that and just not do the gatekeeping / condescension that the other providers are starting, and become more mainstream.
Even without the politics, Elon has shown that he will weaponize his platforms against people/companies he personally doesn't like (e.g. specific bans/demotions to external sites like Substack and Bluesky).
Using Grok is therefore a supply chain risk and it's not nearly good enough to offset that risk.
I do just want to focus on the 'even without the politics' asterisk though because sometimes there is a risk people think everyone on x side (x meaning 'a given side', not x.com) is wrong
You can claim Elon bought x as some sort of power trip. Fine. Willing to entertain it, I have no dog in the fight. I'm not a member of the Elon fan club. And yet Twitter (under Dorsey though I don't think he was involved) was banning tons of people under guises of 'misinfo' that wasn't misinfo
Americans are 4% of the world's population, and even among those 4% at least half don't give a shit. the rest of us give even less of a shit, we don't have the luxury to be principled.
You can very roughly proxy popularity of close-sourced models through OpenRouter token throughput. Grok has an order of magnitude less OpenRouter usage than Claude, GPT, even Gemini.
They were missing a harness like Claude Code or Codex (terminal). However they recently released Grok Build, which is probably the fasted I've used, in terms of responsiveness, but didn't have a model at Opus 4.7/8 level. The thing is if they add 4.5 to Grok Build and keep improving the harness I think it can compete (cheaper and faster).
I've been using Grok Build over the last couple weeks. It's actually a very good CLI. The Grok Build 0.1 model isn't great but can also use Composer 2.5 which is excellent. Well worth trying.
If they were a frontier lab, you'd know.
still waiting for a proper gui for grok build
terminal is nice but codex desktop app is very useful
Just as a reminder, Musk's actions lead directly to the deaths of thousands of people in the third world: https://archive.ph/20250629012329/https://www.nytimes.com/20..., https://www.propublica.org/article/kenya-trump-usaid-world-f...
Enjoy your chatbot!
Not available for Europeans yet. :(
Luckily, basic VPN to US is enough to use it. Just tested it.
I think it should be available through Cursor?
EDIT: Tested myself, it's actually NOT available from EU. But with a Swiss VPN it works :)
We will probably see it when it's available for everyone.
This is the first time I see a lab region locking a model though.
> This is the first time I see a lab region locking a model though.
I think Facebook/Meta was first with this, can't remember exactly what model release but one/some of them had terms locking out EU/EEA residents from using it/some specific features of it.
xAI is under criminal investigation in the EU
Who isn't
Contrary to what you're implying, that's more of a reflection on the typical US corporation than the EU.
Hilarious comment now when they try to push Chat Control again.
I'm not - then again I didn't launch a image generation model advertised as having a spicy mode so that might have something to do with the coincidence.
They say "EU availability is expected in mid-July". So next week or so.
Another subpar model. Why don't they go open weight?
Isn't this the same Twitter company that was supposed to go bankrupt a few years ago? Now it is somehow part of a Space company that has an AI division inside of it?
I think we are going to be waiting a long time for Twitter / X to go bankrupt as it was (erroneously) predicted a long time ago.
Twitter was supposed to go bankrupt if you only read news articles from journalists about it. If you looked at Musk's operating track record, you might have had a different opinion.
In the transaction announcement (xAI buying twitter) twitter reported $12b in debt on acquisition, roughly the amount originally sourced ($13b), so it apparently made good on its debt covenants during the operating period. I have no idea if it received additional capitalization from Musk to do that or not.
That said, the deal was classic Musk - anybody who went on the equity ride with him in Twitter just KILLLED it; xAI was valued at $80bn and twitter at $33bn, so the owners there became 30% owners of xAI. xAI was acquired for $250bn at a SpaceX valuation of $1 trillion, or 20% of the resulting entity, so the twitter stock was 6% of spaceX at about $2 trillion, or $120bn on an equity purchase price basis of $30bn. and that $120bn in value is on really good daily trading volumes; lots of depth.
That was the point of the bailout. Twitter is already a rounding error so no one will notice if it goes to zero.
Don't think it was going to zero anyway. They only had to worry about servicing their debt, they were doing well other than that. And even then they were probably fine.
I am not certain what financials you were looking at but Twitter was unable to ever meet the debt servicing costs for the leveraged buyout alone. It also had overhead costs and other debts that were entirely out of scope for being covered.
The latest Bloomberg reports before the xai merger/acquisition showed a dramatic turn around in their financial situation.
In the beginning it wasn’t good but they would have been fine after that. There are no credible reports to the contrary
twitter was "acquired" by xAI which was then "acquired" by SpaceX as part of the IPO strategy, (and part of a strategy of giving the investors on the hook for the twitter acquisition a return). Who knows how it performs, but yeah, now that it's the social media arm of the SpaceX conglomerate, it will likely be around for a long time, especially since it serves the basic function of stroking Musk's ego.
It is right and proper to view twitter as a loss leader propaganda arm.
None of them go bankrupt. The whole thing will just get stuffed into a larger Matryoshka egg that IPOs for eleventy trillion dollars in 10 years.
I'd say the prediction is correct, as the acquisition is more or less just a better way to capitalize on the bankruptcy.