Kinda funny that their "cost-vs-performance" chart looks the same as the one for Composer 2.5[1], except that it includes Composer 2.5 at a completely different spot.
What are the chances that CursorBench ranks Cursor's model highest, and Cognition's bench ranks Cognition's model highest? Both are to be RL'd from Kimi as a base model, BTW.
I'd posit that it's not deliberate deception, but for both companies their training data and benchmarks come from the same dataset (Devin/Cursor interaction logs) so they naturally overfit.
I actually started typing the same point that the chances are actually high because of train/eval overlap then realised you answered your own question with that same observation.
It is interesting though!
Perhaps in some way this means we should decide which eval set aligns best with our taste?
Back to the blog post. This is an excellent write up of an excellent technical achievement.
I have a lot of respect for the Cognition/Devin (always "Windsurf" to me) and Cursor teams.
I found it interesting - but justified - that they referred to themselves as a foundation lab rather than a dev tools company.
Apparently 'free' on the $20/mo Devin plan (presumably within some quota still)
and that is "via Cerebras at 1000 TPS" according to the announcement
I live on Opus 4.8 High and their benchmark scores SWE-1.7 slightly higher ... if at all realistic that sounds like a great deal ... too good to be true?
We need more models that optimize for coding and that can be cheaper than frontier models, like what SWE 1.7 and composer 2.5 are trying to do. I don't think there's an effort to make something GLM-5.2 level but focused only on coding.
Defining what "coding" means now, and how quickly we fall off the capability cliff seems increasingly important.
Today my "coding" sessions often enough begin with real life problems, where I discuss domain or inter-domain things, ranging from business, economics, psychology, etc. Being able to do all of that with one model is something I am willing to pay a premium for.
Of course not having to pay the premium, because the routing is smart or whatever, would work just the same for me. I just would not want to have to think about it.
OpenAI do have codex models, which are half the price. I haven't used them enough to comment on the quality though.
I remember them saying a few years ago that, they didn't think it was worth specializing models for code, because their general purpose models kept beating them. I guess they changed their mind? Since they did start making codex models again.
I use this model. It's pretty good but not Opus 4.8 or Fable levels obviously. I'm really hoping we get more models like it (and better) soon. I run it locally and it's great that way.
Qwen3-coder-next is very usable. But I don't think it's as good as Qwen3.6-27B (though it does run faster on my hardware). It would be great if we could get a Qwen3.7-coder, but I'm not going to hold my breath.
While I am skeptical of the results here, I am very excited for this new trend of making models faster. Running capable models at 1k TPS is more valuable for me than running better models at 30 TPS. I can only imagine the trend continues to move from "let's only make models smarter" to just incremental intelligence gains but with step improvements in speed.
Why? I'm personally on the opposite end. Less babysitting/higher quality means more time goes back to me/the user. 1000tps of bad code means you have to keep validating the output and circling back.
Indeed. For me opus 4.8 is good enough. If only it would be 100 times faster. You could run it in self verification loops much much faster. It sometimes takes 15 minutes for me to complete a simple task. For example configuring AWS agentcore and deploying an agent on it. Takes forever with Claude with constant issues it tries to solve.
Yes, and not only that but you can't even access it via API, you can only use it in Devin (formerly Windsurf).
I'm an OpenCode user, but I'll fall back to Claude Code if I want to use Opus end to end for something, given my company has a subscription. But I'm not using yet another tool and subscription for a model that isn't even winning.
A company whose first demo was completely fraudulent announces that its model beats GPT-5.5, on its own benchmark? I’m gonna wait a little before I trust this.
This whole company seems to optimize for raising money and impressing VCs. Lying about their products, ignoring consumer market to target enterprise, bragging about how they work their employees like slaves, and writing these posts full of intimidating technical jargon...
To be fair it does seem like most AI startups are now like this (particularly when it comes to constantly mentioning how hard they work and ignoring consumer markets).
Is it just me or does all that* seem pretty tame by today’s standards? Not saying it’s right, but it barely raises eyebrows. Sounds like a pretty typical startup demo.
* Based on the first comment in the link that claims to summarize the video.
Because papers are often referred to by the first author’s name, and often the first author is the primary researcher and therefore deserves the extra credit. When two or more primary authors are equally involved, they’ll often do a random ordering but annotate this so that no one thinks one did more than the others.
Imagine how far community might have pushed if 2 past versions of 'morally superior' Anthropic and 'completely Open AI' open sourced their models for the community to build on top of them
Should as in "would it be nice?" - yeah. Should as in they have to? No.
> Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the “Software”), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so
You can do pretty much anything you want with an MIT license.
There is no obligation to do that. I think the landscape would be very different now if one of the big labs had released an earlier “frontier” model under copyleft that requires sharing fine tunes. I hope it still happens.
I'm looking forward to trying this out. I've been using SWE 1.6 quite a lot for grunt work alongside Opus for higher level planning and tricky stuff - a good combo.
As a (former) Windsurf user I'm pretty happy with the progress of the Cognition/Devin ecosystem after they took over Windsurf, now known as Devin Desktop.
The reality is most people building their own models and providing that alongside SOTA ones don't really care about how great these models are. They just prove that 'hey we are smart enough to build our own models so you can trust us instead of going with a single provider like Claude via Claude Code', also a cheap alternative for cost sensitive/free users - at least this was the case for Windsurf, not sure if Devin Desktop still has that tier. They just need to hillclimb the benchmarks and show something reasonable enough there.
At work I wouldn't want to use anything else. Compared to my salary a Claude subscription (or two) is cheap
For hobby projects I've completely switched to DeepSeek v4 pro. I spend less than on a $10 Claude plan and am not subjected to quota limits (when I have time and motivation, the last thing I want is a 5 hour quota running out). And the difference in model performance is fine for those smaller projects, most of which will end up abandoned or in a state of "good enough" anyways
And for utility tasks, those 30b models are also great.
I'm a big fan of gemma4
Not true since a few months, genuinely try GLM 5.2 and Minimax M3, especially in adversarial/gating... as a general model, I can agree, but as a coding model, they are not bad, comparable to maybe Opus 4.5 in real usage which is quite impressive.
I use GLM or DS4 to help me draft a better initial prompt with more information that I then give to Sonnet 5/Fable/GPT5.5. While benchmarks show the open models close to frontier level, my experience with them is drastically different. I have high confidence that Fable or GPT will 1 shot solutions.
At least with low level programming languages. They're all very good for webdev stuff.
I think showing the API prices for competitors that people don't really pay for that way is all that useful. I do like that it's provisioned by Cerebras though. I think I'd have leant towards focusing on the TPS.
I've always had mixed feelings about Cognition. Obviously they have some very, very smart people working there (I even know a few), and they do make real products. But at the same time, they've made suspicious marketing claims more than once and even been caught making outright fabricated ones; and while they certainly seem to have shaped up from that, I still find their claims to be in a sort of grey area where they seem to avoid unfavorable comparisons and lean on their own benchmarks. Certainly when I've tried their models they have not been nearly as useful as comparable versions of Claude, GLM, etc. -- though I haven't had a chance to try SWE-1.7 yet.
Kinda funny that their "cost-vs-performance" chart looks the same as the one for Composer 2.5[1], except that it includes Composer 2.5 at a completely different spot.
What are the chances that CursorBench ranks Cursor's model highest, and Cognition's bench ranks Cognition's model highest? Both are to be RL'd from Kimi as a base model, BTW.
I'd posit that it's not deliberate deception, but for both companies their training data and benchmarks come from the same dataset (Devin/Cursor interaction logs) so they naturally overfit.
1. https://cursor.com/blog/composer-2-5
Good observation.
I actually started typing the same point that the chances are actually high because of train/eval overlap then realised you answered your own question with that same observation.
It is interesting though!
Perhaps in some way this means we should decide which eval set aligns best with our taste?
Back to the blog post. This is an excellent write up of an excellent technical achievement.
I have a lot of respect for the Cognition/Devin (always "Windsurf" to me) and Cursor teams.
I found it interesting - but justified - that they referred to themselves as a foundation lab rather than a dev tools company.
I think it's also telling that they left out the usual hallmarks of the Pareto distribution: GLM 5.2, Qwen 3.7, Minimax M3, and Mimo 2.5
https://arena.ai/leaderboard/code/webdev/pareto
> they left out ... GLM 5.2
They did not.
Okay, let's give software engineers a break for a bit and focus on obsoleting other high-linguistic context occupations.
Like diplomacy? heh
https://devin.ai/pricing
Apparently 'free' on the $20/mo Devin plan (presumably within some quota still)
and that is "via Cerebras at 1000 TPS" according to the announcement
I live on Opus 4.8 High and their benchmark scores SWE-1.7 slightly higher ... if at all realistic that sounds like a great deal ... too good to be true?
We need more models that optimize for coding and that can be cheaper than frontier models, like what SWE 1.7 and composer 2.5 are trying to do. I don't think there's an effort to make something GLM-5.2 level but focused only on coding.
Defining what "coding" means now, and how quickly we fall off the capability cliff seems increasingly important.
Today my "coding" sessions often enough begin with real life problems, where I discuss domain or inter-domain things, ranging from business, economics, psychology, etc. Being able to do all of that with one model is something I am willing to pay a premium for.
Of course not having to pay the premium, because the routing is smart or whatever, would work just the same for me. I just would not want to have to think about it.
Qwen was doing something like this with their coder models. But alas, they seem not to be releasing those anymore. Last one was Qwen3-coder-next.
Its crazy that OpenAI and Anthropic themselves aren't doing that. No attempts at reducing inference cost for code as far as I know from them.
OpenAI do have codex models, which are half the price. I haven't used them enough to comment on the quality though.
I remember them saying a few years ago that, they didn't think it was worth specializing models for code, because their general purpose models kept beating them. I guess they changed their mind? Since they did start making codex models again.
They also stopped making them again.
I use this model. It's pretty good but not Opus 4.8 or Fable levels obviously. I'm really hoping we get more models like it (and better) soon. I run it locally and it's great that way.
Qwen3-coder-next is very usable. But I don't think it's as good as Qwen3.6-27B (though it does run faster on my hardware). It would be great if we could get a Qwen3.7-coder, but I'm not going to hold my breath.
Wait Devin has a CLI?
Time to support it in my agent IDE just like Cursor's...
While I am skeptical of the results here, I am very excited for this new trend of making models faster. Running capable models at 1k TPS is more valuable for me than running better models at 30 TPS. I can only imagine the trend continues to move from "let's only make models smarter" to just incremental intelligence gains but with step improvements in speed.
Why? I'm personally on the opposite end. Less babysitting/higher quality means more time goes back to me/the user. 1000tps of bad code means you have to keep validating the output and circling back.
Indeed. For me opus 4.8 is good enough. If only it would be 100 times faster. You could run it in self verification loops much much faster. It sometimes takes 15 minutes for me to complete a simple task. For example configuring AWS agentcore and deploying an agent on it. Takes forever with Claude with constant issues it tries to solve.
Not finding anything about this while searching huggingface: https://huggingface.co/search/full-text?q=SWE-1.7 i assume this is another closed source model?
Open weight models should have GPL-like license where it says if you train model on it, it needs to be open weight as well.
Yes, and not only that but you can't even access it via API, you can only use it in Devin (formerly Windsurf).
I'm an OpenCode user, but I'll fall back to Claude Code if I want to use Opus end to end for something, given my company has a subscription. But I'm not using yet another tool and subscription for a model that isn't even winning.
Would have been worth a consideration if it could have been used beyond it's own harness. Unfortunately, doesn't seem to be the case.
https://x.com/theodormarcu/status/2074896486047834380
Oh, that changes everything. If I wanted an arranged marriage I could go back to Claude Code.
A company whose first demo was completely fraudulent announces that its model beats GPT-5.5, on its own benchmark? I’m gonna wait a little before I trust this.
This whole company seems to optimize for raising money and impressing VCs. Lying about their products, ignoring consumer market to target enterprise, bragging about how they work their employees like slaves, and writing these posts full of intimidating technical jargon...
To be fair it does seem like most AI startups are now like this (particularly when it comes to constantly mentioning how hard they work and ignoring consumer markets).
> it does seem like most AI startups are now like this
Remember when AGI was going to replace all jobs in 6 months? It's always been like that.
Link for this?
https://news.ycombinator.com/item?id=40008109
Is it just me or does all that* seem pretty tame by today’s standards? Not saying it’s right, but it barely raises eyebrows. Sounds like a pretty typical startup demo.
* Based on the first comment in the link that claims to summarize the video.
> "A company whose first demo was completely fraudulent"
Could you expand on this?
Would love to see these companies use benchmarks done by third parties.
they are right there? it shows swe-bench multilingual and terminal bench
What happened ?
Unrelated: what's the point of "*equal contribution"? Why would someone specify this
Because papers are often referred to by the first author’s name, and often the first author is the primary researcher and therefore deserves the extra credit. When two or more primary authors are equally involved, they’ll often do a random ordering but annotate this so that no one thinks one did more than the others.
Interesting. Thank you
Feels like they discovers that if you build your own benchmark, you can win it
Pretty sure most benchmarks are being gamed by people training on the test set deliberately or accidentally anyway.
Open source for the win!
Imagine how far community might have pushed if 2 past versions of 'morally superior' Anthropic and 'completely Open AI' open sourced their models for the community to build on top of them
Is this open source? I can't find a link to download the weights.
It's based on an open weight model (Kimi 2.7) so shouldn't it also be open weight?
> so shouldn't it also be open weight?
Should as in "would it be nice?" - yeah. Should as in they have to? No.
> Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so
You can do pretty much anything you want with an MIT license.
There is no obligation to do that. I think the landscape would be very different now if one of the big labs had released an earlier “frontier” model under copyleft that requires sharing fine tunes. I hope it still happens.
Not open source. Also, not available beyond it's own harness.
Heads up to anyone else curious, I installed the Devin CLI and SWE-1.7 is not currently available there.
I'm looking forward to trying this out. I've been using SWE 1.6 quite a lot for grunt work alongside Opus for higher level planning and tricky stuff - a good combo.
As a (former) Windsurf user I'm pretty happy with the progress of the Cognition/Devin ecosystem after they took over Windsurf, now known as Devin Desktop.
These models are never as good, the benchmarks dont tell the full story
The reality is most people building their own models and providing that alongside SOTA ones don't really care about how great these models are. They just prove that 'hey we are smart enough to build our own models so you can trust us instead of going with a single provider like Claude via Claude Code', also a cheap alternative for cost sensitive/free users - at least this was the case for Windsurf, not sure if Devin Desktop still has that tier. They just need to hillclimb the benchmarks and show something reasonable enough there.
Benchmarks are just vibes with error bars... wake me up when it survives a week on a real codebase without hallucinating a package that doesn't exist.
Funny, the cheerleading at HN for leading Chinese models, but a non Chinese lab (building on top of a Chinese model) gets dissed here.
It's simple: close weights = not welcome.
It's almost as if HN users aren't all the same.
all the open source models are a waste of time relative to the bleeding edge from openai/anthropic
At work I wouldn't want to use anything else. Compared to my salary a Claude subscription (or two) is cheap
For hobby projects I've completely switched to DeepSeek v4 pro. I spend less than on a $10 Claude plan and am not subjected to quota limits (when I have time and motivation, the last thing I want is a 5 hour quota running out). And the difference in model performance is fine for those smaller projects, most of which will end up abandoned or in a state of "good enough" anyways
And for utility tasks, those 30b models are also great. I'm a big fan of gemma4
ive just got better things to do with my life than fuss with an inferior model. its like why hire a dumb employee over a smart one
I think you misspelled "I've got plenty of money".
Not true since a few months, genuinely try GLM 5.2 and Minimax M3, especially in adversarial/gating... as a general model, I can agree, but as a coding model, they are not bad, comparable to maybe Opus 4.5 in real usage which is quite impressive.
I use GLM or DS4 to help me draft a better initial prompt with more information that I then give to Sonnet 5/Fable/GPT5.5. While benchmarks show the open models close to frontier level, my experience with them is drastically different. I have high confidence that Fable or GPT will 1 shot solutions.
At least with low level programming languages. They're all very good for webdev stuff.
yeah but why waste your time on these models, just use the one that gets the better results
I actively prefer GLM-5.2 for some tasks. For simple tasks the results are just as good as e.g. Opus, and it produces results significantly faster.
Because you can get them from more trustworthy providers or with hardware encryption.
i trust anthropic/openai with my data far more than some random startup.
I was going to respond until I saw your account name lol.
haha i outsource my thinking to the smartest model
I think showing the API prices for competitors that people don't really pay for that way is all that useful. I do like that it's provisioned by Cerebras though. I think I'd have leant towards focusing on the TPS.
I've always had mixed feelings about Cognition. Obviously they have some very, very smart people working there (I even know a few), and they do make real products. But at the same time, they've made suspicious marketing claims more than once and even been caught making outright fabricated ones; and while they certainly seem to have shaped up from that, I still find their claims to be in a sort of grey area where they seem to avoid unfavorable comparisons and lean on their own benchmarks. Certainly when I've tried their models they have not been nearly as useful as comparable versions of Claude, GLM, etc. -- though I haven't had a chance to try SWE-1.7 yet.