106 comments

[ 0.30 ms ] story [ 45.9 ms ] thread
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.

> "A company whose first demo was completely fraudulent"

Could you expand on this?

Their product has far evolved beyond this (of course with the large amount of money being poured into it) and is now used by a lot of traditional companies (banks etc). Also the SWE models were originally built by Windsurf and this seems to be on top of that (after acquisition) although the original SWE-1 models weren't that groundbreaking.
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).
I highly respect many people at cognition but yeah that's put a sour taste in my mouth.

I want to work in the AI space on actual AI research, at any part of the stack. Even if I'm developing training infra - as long as people are advancing knowledge of what intelligence could be.

But it seems like either it's big labs or grifters, that's it, and even the big labs, at least publicly, seem very grifty at times. Not like I have the technical chops probably, but still.

This is inevitable when the primary incentive is to raise aggressively. Overall I dont find cognition blogs that jargony, there are definitely worse offenders
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

(comment deleted)
Not open source. Also, not available beyond it's own harness.
These models are never as good, the benchmarks dont tell the full story
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.
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.
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.
Unrelated: what's the point of "*equal contribution"? Why would someone specify this
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.

> Today my "coding" sessions often enough begin with real life problems

intuition is that your sessions consists of 10% of domain related reasoning, and 90% of code plumbing. Those 90% could be moved to cheap and efficient specialized and focused model.

Possibly! It's just hard to reason about from the outside. When does the model benefit from all the ambient knowledge? Idk.

Regardless, it's fairly obvious to me that none of what I do now will require "frontier models" for much longer. Models are getting better more quickly than my problems are getting harder.

But that 10% is the most important part! Getting the plumbing wrong means you might have bugs or your code is brittle. Getting the domain-specific business logic wrong means your product doesn't fundamentally solve the correct problem.
but that is not model problem

most agentic coding app can use powerful model for planning/reasoning then use "budget" model to do ground work

> most agentic coding app can use powerful model for planning/reasoning then use "budget" model to do ground work

I've had terrible success using budget models to do ground work. The justifications that the budget models will use and document, polluting the rest of the session, are sometimes just insane. Like making code compatible with a bug that was implemented within the same session, not handling errors due to precedence in the code it just implemented, etc. I DO have success using the heavy models with lower effort, and using budget models on relatively changes post ground work. But major planning and initial ground work, I just get absolutely slop if I use a budget model.

If you're doing web stuffs, or GUI, then the budget models seem fine.

And make it run locally!
This isn't as easy as it sounds. Every ML model is struggling to balance between generalization and test performance.

Taking a good model like GLM5.2 and just fine tuning it on coding can decrease real world performance due to mechanics like catastrophic forgetting. There is also other interesting behaviors were training on a broad training set can improve coding performance because there is positive transfer.

There is 100% an effort to make solid coding focused models, but it is very hard to do that without including capabilities across a broad set of adjacent tasks.

As you say- LLMs are fundamentally good because of their generalism. Distillation, ablation, ft all tend to be hacky and in some way hurt the model
Heads up to anyone else curious, I installed the Devin CLI and SWE-1.7 is not currently available there.
It is now, at least for me. Latest Devin CLI version: v3000.1.27 I don't think SWE is geo-restricted.
Looks like now it shows up but it's not available to try on a free plan.
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.

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.

agent lab, not foundation lab
Agreed on the likely mechanism. I'm not sure "overfitting" is even the right description. These things are of course absurdly complicated, and evaluating their quality down to a single number involves a lot of judgement and trade-offs. I think it's more "you get what you measure" which is true in human organizations too. Define a KPI and people work hard to make it go up, even if it's not quite right or has bad side-effects.
At this point I barely put any value in any of the benchmarks. I just use the models for coding (and related things like software product design/planning/ideation/etc.) tasks and judge them subjectively, and also see how others judge them subjectively on HN and Twitter.
I use benchmarks…

…that are my own private internal suite on my own code bases where I can judge the output properly

I also measure wall clock time to completion which has been a surprising separator in practice.

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.

Okay, let's give software engineers a break for a bit and focus on obsoleting other high-linguistic context occupations.
And to do that you’ll need development so until we’re all out of a job they’ll keep pushing. Once automating is automated it’s done.
Any day now... Just a bit more space in the context window, trust me bro.
We had decades of "just more and faster bits" in computing and that produced a lot of new capabilities.
software and agentic workflows will obsolete those things

RL environments building on top of each other will get these models there

needs people doing software development lifecycles to figure it out and implement

Feels like they discovers that if you build your own benchmark, you can win it
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.
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.
Harness-wrapper tools that support multiple harnesses and allow sharing workspace features (skills, slash commands, etc.) between them will be meta.
Ironically, Devin Desktop is one of those tools. It supports any harness that supports ACP (which is most of them)—you can use Claude Code, Codex, OpenCode, etc from the Devin Desktop UI.

I'm currently experimenting with OpenSpec[0] as the "framework" and using different subscriptions for different parts of the spec-driven process: Opus via Claude Code for exploration, Devin SWE for building, and GLM 5.2 via the Z.ai Coding Plan for verification. I don't love having to mix and match harnesses, but in practice it's barely more effort than switching models.

[0]: https://openspec.dev/

Is that surprising? It is standard "embrace, extend, extinguish" from a company not in a strong enough position to do the third one.
I really don't want harness lock-in. I am trying to decouple myself from Claude Code now. I love the model of OpenRouter and being able to switch models at will let's your harness focus on your personal tooling and you can easily switch to the flavor of the month LLM with a single slash command instead of rewiring your entire workflow to use a harness to use a model.

I like Cerabras, but I really wish they would make more of their hosted models generally available.

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.
id rather iterate multiple times than wait 15 minutes to notice it made a mistake.
Again, my point is exactly the opposite. Higher quality implies a mistake isn't made in a significant % of cases.
It's a lossy conversion though. "Mistake" is relative to the stated goals and specifications which are often heavily lacking. So unless you write with a high degree of architectural and implementation specificity then it might make very high quality code that is still not what you wanted.
You can ask for a complete feature/app/business. Or you can split up the work into verifiable/testable pieces and rely on a high quality AI to deliver. As time goes by the pieces will get larger as capability grows. I still trust myself and my experience when arch is involved, but AI has been great at tackling lower level stuff. And with Fable I don't really care it takes a while for it to complete, as I know I can trust it a lot more (which is what I personally prefer). Yes, with a 10k tps model you can iterate quickly. But that's not me personally.
But some requirements you don't realize you have until you start building. With a fast model, you can surface those really quickly and have more time to iterate and explore different solutions. With a slower but smarter model, you just hope that what it produces after an hour is what you were imagining.

And yes, with Fable, the chance of that is higher than with SWE/Composer, but in my experience it's not so much higher that the extra time and cost is worth it. But it certainly depends on your goals and what you're building.

So i agree with you, but there's no SOTA model that i don't have to babysit. I'm not going to just throw a large pile of code in there unreviewed, and so what i want is faster iteration on code in logical, reviewable chunks. Ie just like i'd normally write myself; small, logical commits.

Faster iteration means i mentally checkout less and am more involved with the code being created.

My hope is that in the far far future, we can get LLMs so fast that i can work in my IDE like normal and the LLM will just be an extension of autocomplete. I can state a goal, rough out functions, code, etc, and it'll just work around me like a very fast pair programmer / autocomplete.

The chat interface is an intermediate step that frankly i hate. The faster it is the less i wait.

Now for vibe-slop i'm making on the side, yea i don't care about speed. But that's not something i'm employed to do or anything i truly care about. It's a different workflow entirely.

I get it, you just prefer to do things differently

> Faster iteration means i mentally checkout less and am more involved with the code being created.

This is a good point I didn't consider and you're right. More interaction brings you closer to the code.

I still think that this is the opposite of what I personally want. Either I write the code (or a large majority of it), and be fully involved; or be more disconnected but more free to focus on other things. The middle ground removes me from the equation, but also requires me to babysit.

High tps is good for deeper agent thinking loops and openclaw etc. I was running cerebus recently doing some data heavy tasks, it managed to crash the server I was submitting posts to. 6 hour task down to ~1hr
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.
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?

The 1000 TPS shows for me as "SWE 1.7 Lightning" and took 14% of my daily quota in one prompt on the $20 plan.

But the normal speed one seems to be free or with very generous limits.

The "Fast" (Cerebras) variant isn't free, only the regular one, which runs closer to 50-100 TPS in my experience with SWE 1.6.
Oh, the free plan said "Slow" so I thought maybe the others had the fast version :)
I used SWE-1.5 and 1.6 when it was Windsurf (before Devin Desktop), it's not that bad (grunt work, tests, can actually plan and implement some medium level stuff) but you get a much much better value and better models (GPT-5.4^) going with a Codex subscription (plus you get resets).

That company truly subsidized its user base to the extreme before, the $15/mo subscription was the best value on Earth paired with weekly deals reducing credits for premium models. Now it's barely any messages for paid models, completely watered down.

FWIW, Cognition has all the Sonnet/Opus/Fable models, and all the GPT ones, as well as GLM, Kimi, and Gemini.
But they don't appear to subsidize them to the same degree. I've only been using Devin for less than a month, but I've been hitting the limits of the $20/month plan way more quickly than I'd expect, and definitely more quickly than with Claude Code or Codex.

So far, Cursor provides the best value for their subscription, but I have to imagine they're basically lighting money on fire. There's no way their current pricing is sustainable.

It's not very clear on the website, couldn't find a list of models or quota/pricing per model anywhere
Wait Devin has a CLI?

Time to support it in my agent IDE just like Cursor's...

How do i use it from opencode/openrouter?!
On artificialanalysis.ai, Kimi 2.7 Code is way worse than GLM 5.2 at everything (general intelligence, coding, agentic tasks).

But here, both Kimi 2.7 and its derivative SWE-1.7 are ahead of GLM 5.2. This tells me the benchmarks they use are cherry-picked.

> This tells me the benchmarks they use are cherry-picked.

Which benchmarks would you have chosen instead, and why?

It honestly seems like there's not a great way to currently benchmark AI.

The ideal way to run these benchmarks would be to give a 3rd party the model to run in an isolated environment so the prompts don't make their way back to the AI engineers.

That seems doable for open weight models, but not for private models.

If you've got money to burn on tokens, the way that seems best to me is to set up a repeatable harness - docker container with a specific past commit from your own project, set of known issues/features that you've already fixed/completed of varying levels of complexity.

Set up a script that launches the harness for each model, prompts them to implement one of the tasks, let it churn until either tests pass or it hits some budget limit.

Then, most importantly, read the transcript and output and judge subjectively - I don't think this actually can be narrowed down to a score, although tokens burned to fix, whether it actually got the tests green etc are all good signals.

(I've done this, but so far only on a codebase that was too complicated with models that were too weak because I didn't want to spend more than a few dollars - results were inconclusive, planning on iterating on my personal benchmark in future)

Likely benchmaxxed. You see it in Qwen and other smaller models all the time.
Composer 2.5 is worse than both; I use it all day for simple stuff. It's Kimi 2.6 in a new outfit.
Cognition... oh what a ride... We were customers when they acquired Windsurf, stopped offering customer support, raised prices, dismantled the brand, and raised prices again. We are not customers anymore. Benchmarks are not the only thing to worry about when you are using models.
I’ve unfortunately had to temper my excitement with Cognition’s models/products given the amount of unwarranted hype they created with Devin on first release, but hopefully this is good.
I work with Devin daily (as well as Claude and a few others) and I can attest it's not a cheap product but it's a good one and it saves me a bunch of time.
Yea I cant convince myself to pay api pricing unless my employer is doing it
(comment deleted)