If you're looking for reason to be skeptical, look no further than the massive delta between the Terminal Bench 2.1 (92.8%) and the Terminal Bench 4 score (27.3%).
Terminal Bench 4 was released a couple weeks ago, so the difference you're seeing between the two scores can be interpreted as "how well does this model generalize to new problems"? More crudely: "how benchmaxxed is this model?"
This is a groundless criticism. TB2.1 is saturated. TB4 is not. Sol xhigh is 90% on TB2.1 but 37% on TB4. Is it also "benchmaxxed"?
Your assumption is that the benchmarks are essentially identical in difficulty, with the only difference being their age and thus whether they could have been trained on.
> Sol xhigh is 90% on TB2.1 but 37% on TB4. Is it also "benchmaxxed"?
A model that was released a couple months ago scores 50% higher than SWE-2, a model released today, on an out-of-sample benchmark. Can I say I’ve come out of this more impressed with Sol?
Like you said, TB2 is saturated. Nobody would bat an eyelash at 90%. And yet here comes SWE-2 coming off top rope with an emphatic 92.4%. this is the definition of bench maxxing.
Benchmaxxing is the default case, and always has been.
It's really, really difficult to avoid it even when you care to stop yourself; and it's not even just a problem in machine learning, it's the standard failure mode of all minds capable of learning, human, animal, artificial.
Even pure genetics has this problem. Viruses and cancers also demonstrate this behaviour, with the bench being evolution's only option: reproductive success.
I take it to mean the benchmarks are a marketing line item, as in, to sell this fucking thing you have to go out there and lie and the way everyone is lying is by doing exactly that, lying. They build for benchmarks and build benchmarks for builds.
You want to make money or not , motherfucker? That’s the game. If you have to literally concoct a fabricated bullshit story about how your model hacked its own computer, then go fucking do it. Trillions. Trillions of dollars is what they want, and to sit and think anything other than human nature is at work here can only be possible in the realm of truly delusional people. It’s a dirty world.
Anyways, the other takeaway is that they are having to LIE to make money on models which means commodification has already occurred and we’re in an entirely new phase.
Is Sol benchmaxxed? Of course it is. Altman was caught in previous attempts trying to game benchmarks, does anyone believe that he's found his moral compass and decided to stop exploiting as much as he can get away with?
I mean, still benchmaxxed, I happen to not consider that a problem
These firms are literally hiring professionals from all fields to teach procedure
To teach processes that can subsequently be done agentically or in automated chains
Its basically infinite permutations of tool calling, except the tools aren't external, they’re baked in upon birth
So yeah still makes sense that the new benchmark has a low score and the older one has a high score. And sure, one day we wont have to debate it and a new model will ace everything. Do you actually want that day to be today?
GLM 5.3 looks strange, because of this Chinese labs benchmaxx moto. So rather they have emergent abilities or...
Also a lot of questions to benchmark because opus 5 is completely useless model right now.
I think that the main problem with opus that they try to solve context size optimization problem, and that is the main reason why it speaks like alien with only one technical dictionary at hand. So why it is so good?
Your post made me wonder if Artificial Analysis had finally moved to TB4 and lo and behold they have and Astra is tied with Fable 5.1 at 53.
That then made me realize that they lower the bars of tied scores so on the site it looks like Astra in second place. Weird. Anyway, yes, so many of these composite benchmark sites are irrelevant if they're not trimming the fat and sticking to the most up-to-date variants.
Yeah, this echoes my thoughts. I will be very surprised if a model with 2.8T parameters reaches the intelligence and capabilities of 10T parameter models. RL can take things far, but not that far.
Closed weights AND benchmaxxed. Somehow this company raised 2bil at a 48bil valuation. Pure insanity. I feel bad for their investors (not really, but... Still). Andreessen Horowitz is being played like a fiddle.
Cognition, the same company that a few years ago demoed a coding bot purporting to be able to autonomously complete upwork tasks, but upon closer inspection was going off the rails and not even completing what was asked?
As others have mentioned this is post trained from Kimi k3, which is already quite capable, so it can't be that bad, but any claimed improvements in performance should be taken with a grain of salt.
A friend recently pushed me to try out their coding platform, Devin, after I decided to move away from Cursor. I had the same reaction: "What, the con artists from like 2024?" But after some cajoling, I gave it a shot and was pleasantly surprised. I guess they learned their lessons, grew up, and are doing good work now, maybe?
Devin CLI is easily the worst-in-class coding harness I've ever been subjected to using, rife with bugs (up to and including dropping answers to the question tool), so I can't say I'm exactly inspired to try anything else the company produces.
Not all that surprising. The original Devin really was just an early attempt at agentic coding before models were really even trained for it. Now that it's a well established pattern and we've figured out what works, I'm not surprised they've morphed into something reasonble.
When they launched Devin it was supposedly at the performance of an engineering intern. Friends who used it found the bad parts of an intern (tons of handholding, review required) but it didn’t learn from mistakes or add throughput.
I think the evolution of the harness and ability to preserve loop context outside the context window has made running these kinds of agentic experiences easier.
sorry so many buzzwords to say, the capabilities to do this kind of work are more accessible and easier to manage, so now it works!
Good to see, and agree they were severely overhyping their product back then.
As others have mentioned, it's really matured a lot and at this point is one of the best cloud-hosted, team-managed coding agents, when factoring overall UX, testing and QA lifecycle via its sandboxes, and its ability to be controlled with an API. We use it quite heavily.
It's coming from a different starting place than Claude Code or Codex are as individually controlled single-developer tools. Devin has been more persistent in pursuing the direction of something that operates more autonomously at the team level, as a peer. And while it might be slightly behind in raw harness ability (maybe?) it's probably ahead on the team-focus.
Where are the model stats? Is this open-weights? If not, why would I use this over DeepSeek Flash 4.1?
I think these competing labs need to realize that no one wants another closed-weight model provider... We aren't even happy with the two we have right now, and their days are entirely numbered. If DeepSeek 4.1 flash is really as good as it's benching, we're probably a month away from 1/3rd of users moving off the closed-weight models in favor of something they have more control over (or is cheaper).
The big labs love to release their new model and quantize after the first week. You don't have that problem using dirt cheap API rates on OpenRouter. DS 4.1 flash is also faster than fast mode Astra. OAI's subscription rates are good value, but now these new open-weight models are nearly as cheap on API usage rates. I honestly can't wait for the day we're not beholden to the two big labs anymore. No wonder there's so much fear pumping happening at the moment from Anthropic and their funded NGOs.
This really just exists so cognition can stop spending API tokens with Anthropic or OpenAI.
Basically any successful AI based service will do this because at scale the frontier models are expensive and you’ll have enough data to fine tune your own.
Same reason Harvey is doing models now and basically every other provider
> If DeepSeek 4.1 flash is really as good as it's benching, we're probably a month away from 1/3rd of users moving off the closed-weight models in favor of something they have more control over (or is cheaper).
Only in the world where the incumbents don't react. Eg if they saw lots of users moving away, they'd drop prices or do something else.
1M cached tokens on deepseek is $0.006, the big labs can't sell anywhere close to this, they have funders expecting returns and huge overhead.
btw I've had a ton of fun with the new deepseek today, I was waiting for my OpenAI 5h limit reset and decided to give it some problems for fun, got pretty great results. Tried some harder problems and still got great results. I don't expect it to be Sol class or anything but I really didn't expect it to be anywhere near this good so we'll see where it ends up. And it's really fun throwing crazy amount of tokens at the wall for ~free instead of watching the subscription limits tick closer while your agents churn away.
OpenAI just paused new subscriptions to their $200 plan. They are in a rock and a hard place. Obviously the Astras and Fables of the world are exponentially more expensive, but for...less than exponential returns. The question is whether they can leverage the marginal advantage into something that justifies the diminishing returns before the bottom catches up to them.
On the one hand you, if you bought a lot of compute a couple years ago (perceived demand, perceived shortage) you are in a good spot temporarily. But the counter to that is that everyone else is becoming more compute efficient so maybe that advantage isn't what people thought it would be. I can almost, almost run DS4.1 Flash at home. 4 sparks can do it at 200+ tokens per second. I have two Sparks, so I am not in the club. Neither is your average laptop owner or gamer either. But your average HN software engineer can probably easily swing 2 sparks.
> we're probably a month away from 1/3rd of users moving off the closed-weight models in favor of something they have more control over (or is cheaper).
DS 4 Flash requires large amounts of memory to run at reasonable quants (I think a system with 160 GB or so). DS 4.1 Flash is even larger, I think around 250 GB.
Any DS version is dumb when compared (in realworld tasks) to Astra/Opus 5, which means, one would spend thousands of dollars, and still need to rely on cloud services to do jobs that are non trivial.
This made me laugh a bit. I was forced to do an evaluation of their shit product twice due to being backed by the same PE firm; "take a look at it again, it's much better now". It sucked the second time also...
Unfortunately yes. It's awful, other than that they at least support both open-weight and proprietary models to route to. So I guess their model router is "fine", but the harness? As I said in a sibling thread on this page:
> Devin CLI is easily the worst-in-class coding harness I've ever been subjected to using, rife with bugs (up to and including dropping answers to the question tool), so I can't say I'm exactly inspired to try anything else the company produces.
I pay for it (mostly because they grandfathered me from the old prices)!
I used to use windsurf as my main editor until they changed their pricing model. Now i use it just to burn my weekly tokens on fable/astra if i remember to that on a task and that's it.
I only used their "DeepWiki" automatic docs, they are pretty decent at getting an overview of a large project and are relatively accurate, with diagrams and anything. Haven't tried out their coding agent stuff.
My company uses it. We have a bunch of seats in an enterprise plan and have been using it for 9 months or so.
It's a great product compared to Copilot. It is also the first AI tool I used heavily outside of creating random images or one off questions.
I'm now using all three, Devin, Claude, Codex. I'm finding Claude and Codex to be much better. One of my biggest gripes is that the web client and desktop client for Devin are two completely different harnesses, so the quality of responses varies greatly.
I use it, and have been happy with it for the most part. Like sibling, I use it for GPT-5.6-Sol and Opus work, and use their free models (GLM 5.2 for the past few months, trying SWE-2 now).
On the one hand I would have expected a completely new model, on the other hand it's an RL-ed K3 go Fable 5 capabilities, which demonstrate that this is probably possible, which is nice.
No one in the US is going to fund pretty good open source with VC money.
US has OpenAI/ Anthropic/ Google/ meta/ SpaceX atleast trying to make frontier foundation models 2 are using VC money + cash flow, last 3 are mainly cash flow + equity and debt.
China has state banks and similar willing to fund lower margin open source labs.
SWE 1.6 was great for small tasks. Very fast and good enough. 1.7 was unusable for me. Took more time thinking than GLM 5.2 and seemed to be generally running in circles. I tried it but abandoned it.
I am skeptical. Lived experience is what matters and I don’t have anyone in my life (Devin shop) saying good things about SWE other than it’s free. Hope I’m wrong and it’s not so bad this time
I think I'm probably in the minority here, but for my line of work, the software engineering and coding is only a small part of the work. I write simulation software, so a deep understanding of physics, math, and how they can be applied to the software is absolutely crucial. I'm assuming this model is tuned to be more focused on SWE topics, and the very reason we seek "multidisciplinary" hires is the also why I actually need a jack-of-all-trades model to back my coding agents.
I'm also in simulation software! Wondering which models you are finding helpful, the models I'm using for general SWE skills are horrible at our simulations and even basic physics/engineering calculation and intuition
I use Claude Opus 4.8 almost exclusively. I have had fairly good experiences with it. One time it derived an entirely novel simulation method different than anything in literature by combining its knowledge about how problems in other fields with similar underlying mathematical structure are solved. That was a bit of a Jacobian Conjecture moment for me.
I'm in the same field and I find GPT to be better at understanding physics conceptually but Claude is better at writing numerical code. I use cursor so many of my sessions start in GPT and switch to Claude.
I like Cognition as a company and hope they succeed. Seemingly excellent engineering org.
I used to really like Windsurf. (Now Devin. Kind of? But also now Antigravity.) I still use it as my editor but haven't touched the agent for a while simply due to the rise of Codex.
Let me grab my calculator and add up the time I have spent reading about model releases since Fable has been released. It seems they all place themselves relative to Fable. I'm sure that time has added up to far greater than 10 seconds. At some point, it ceased to be a meaningful differentiation. This is especially true when I put the model through real usage.
Please correct me if I'm wrong, but this appears to require Devin to use? I'm disappointed to see I need to use a bespoke platform to interact with this agent, to the point that I probably won't be trying it.
But I don't want to use your CLI. I already have my own harnesses and workflows. The friction is too high to "just try out" a new model like this. It would be preferable if I can evaluate it over, say, open router like all the other models and then decide from there if it's worth downloading a bespoke tool chain for only 1 lab's models
I just gave it a try and it doesn't appear to be free, it used up some of my on demand usage. It does say 75% off though. Seems like for Pro subscribers SWE-1.7 is free, maybe SWE-2 is free for them?
If its weights are open, that covers a multitude of other sins. Sufficiently-strong performance on the part of the new model would justify adapting existing tools to work with it.
As an Econ graduate, pretty cool seeing Pareto in the "AI-bro" zeitgeist. Slightly surreal watching a 1906 welfare economics idea get rediscovered as a plotting convention. The original, if anyone fancies 579 pages of Italian: https://archive.org/details/manualedieconomi00pareuoft. There is an English translation somewhere.
Pareto frontiers are pretty commonly invoked to describe tradeoffs in computer science and have been for quite a while. I remember the term being used in one of my early algorithms courses to describe the tradeoff between data structures with fast writes, ones with fast reads and ones that tried to balance the two.
IIRC cognition boasted about hiring a lot of competitive programmers and algorithms experts back when they released Devin, so it tracks that they'd use the term.
It's interesting watching people throw about pareto frontiers sort of like how RF nerds approach the shannon limit (in a practical real world sense of the term, like charting possible modulations/data rates on a two way satellite modem's manufacturer datasheet).
Pareto leaked out of the sociology/econ bubble a long time ago :) Pareto principle, Pareto efficiency, Pareto distribution have been in the pop-sci buzzwords for quite awhile, I probably encountered it first in the 4-Hour Workweek. I don't think you can read a self-help book without the author introducing it as a groundbreaking principle to live your life by.
Yeah, I'd expect model performance to be super spiky on SWE work, at least they admit it with the name of the model. It's distilled from an already-distilled model.
Maybe still worth it if their "64% cheaper" figure holds.
I presume post training is significantly easier than the distillation/training the top Chinese labs are doing.
I wonder if, similar to the American labs, they'll become stingy with their weights once they start getting immediately undercut by a wave of slightly better derived models.
At work I setup a cloud worker, where i can spin up as many concurrent agents I want, with unlimited fable 5.1 (thanks employer!!).
I now just work from my phone, and speak into the agents as they run. I dont write code and I dont write documents. I work on very complicated distributed systems. I dont open my laptop most days. Its a legacy brick I carry around.
Some of my coworkers are still doing things by hand, and are working long hours to produce 25% of the output (when considering hours worked). I stay quiet with my setup. We are in the end times for this job for the people that can see clearly how to automate their own job
that's why after 1 year of product development of these AI 20x maxxed speed, we reached AGI 'wizards', there's really no difference in output, outstanding bugs no longer get solved and sites still suck, even doing things that were just regular development 20 years ago. Are you sure they aren't only producing 2.5% of your output that you manage just by farting into your phone? Are you sure it's 25% really? Seems way to high, days when I have diarrhoea my AI agents move even faster
Do they have anything as good as Amp (which is able to use free models)? Amp has been in the lead for almost a year now and doesn't seem to be relinquishing it.
I'm using Codex, Gemini etc, they all have desktop apps and have a plan, how do i use SWE-2? Thats is a problem they have. I'm not about to switch out my workflow and plans with a shiny LLM that looks benchmaxxed and graph maxxed.
SWE-1.5 was surprisingly good when I used it last. I feel like Cognition is one of the solid players that’s flying a bit under the radar while Anthropic and OpenAI race to IPO.
Odd to get downvotes simply for sharing my experience. Like it or not, Cognition has a good frontier model and they are building serious products. They are working hard which is how you become successful. Sorry if that ruffles your feathers.
all this while Cognition doesn't even make the top page of search results, that's some impressive stealth.[1]
(just to be clear, I am a far-left activist who spends most of my time working on funding Social Security Trust Funds (OASI & DI Solvency), which could impact my search results - this was while I was logged in.)
If you're looking for reason to be skeptical, look no further than the massive delta between the Terminal Bench 2.1 (92.8%) and the Terminal Bench 4 score (27.3%).
Terminal Bench 4 was released a couple weeks ago, so the difference you're seeing between the two scores can be interpreted as "how well does this model generalize to new problems"? More crudely: "how benchmaxxed is this model?"
This is a groundless criticism. TB2.1 is saturated. TB4 is not. Sol xhigh is 90% on TB2.1 but 37% on TB4. Is it also "benchmaxxed"?
Your assumption is that the benchmarks are essentially identical in difficulty, with the only difference being their age and thus whether they could have been trained on.
> Sol xhigh is 90% on TB2.1 but 37% on TB4. Is it also "benchmaxxed"?
A model that was released a couple months ago scores 50% higher than SWE-2, a model released today, on an out-of-sample benchmark. Can I say I’ve come out of this more impressed with Sol?
Like you said, TB2 is saturated. Nobody would bat an eyelash at 90%. And yet here comes SWE-2 coming off top rope with an emphatic 92.4%. this is the definition of bench maxxing.
4 replies →
Benchmaxxing is the default case, and always has been.
It's really, really difficult to avoid it even when you care to stop yourself; and it's not even just a problem in machine learning, it's the standard failure mode of all minds capable of learning, human, animal, artificial.
Even pure genetics has this problem. Viruses and cancers also demonstrate this behaviour, with the bench being evolution's only option: reproductive success.
6 replies →
> Sol xhigh is 90% on TB2.1 but 37% on TB4. Is it also "benchmaxxed"?
Yes.
"Sol xhigh is 90% on TB2.1 but 37% on TB4. Is it also "benchmaxxed"
Yes! extremely sharp RL-fried model. byte perfect hash gates and soak and smoke tests abound.
Fixed benchmarks will be debunked eventually I'd think. Better to use synthetic problems.
Too easy to game the numbers, and too easy to baselessly accuse companies of gaming the numbers, not to mention how you even define that.
1 reply →
> Sol xhigh is 90% on TB2.1 but 37% on TB4. Is it also "benchmaxxed"?
Yes? Just like every single model from every single AI lab.
I take it to mean the benchmarks are a marketing line item, as in, to sell this fucking thing you have to go out there and lie and the way everyone is lying is by doing exactly that, lying. They build for benchmarks and build benchmarks for builds.
You want to make money or not , motherfucker? That’s the game. If you have to literally concoct a fabricated bullshit story about how your model hacked its own computer, then go fucking do it. Trillions. Trillions of dollars is what they want, and to sit and think anything other than human nature is at work here can only be possible in the realm of truly delusional people. It’s a dirty world.
Anyways, the other takeaway is that they are having to LIE to make money on models which means commodification has already occurred and we’re in an entirely new phase.
Is Sol benchmaxxed? Of course it is. Altman was caught in previous attempts trying to game benchmarks, does anyone believe that he's found his moral compass and decided to stop exploiting as much as he can get away with?
16 replies →
I mean, still benchmaxxed, I happen to not consider that a problem
These firms are literally hiring professionals from all fields to teach procedure
To teach processes that can subsequently be done agentically or in automated chains
Its basically infinite permutations of tool calling, except the tools aren't external, they’re baked in upon birth
So yeah still makes sense that the new benchmark has a low score and the older one has a high score. And sure, one day we wont have to debate it and a new model will ace everything. Do you actually want that day to be today?
Those 27.3% are still in the ballpark of modern models:
- Sonnet 5 - 12.4%
- Luna - 17.3%
- Grok 4.6 - 20.3%
- Sol - 37.3%
- GLM 5.3 - 41.8%
- Opus 5 - 51.8%
GLM 5.3 looks strange, because of this Chinese labs benchmaxx moto. So rather they have emergent abilities or...
Also a lot of questions to benchmark because opus 5 is completely useless model right now.
I think that the main problem with opus that they try to solve context size optimization problem, and that is the main reason why it speaks like alien with only one technical dictionary at hand. So why it is so good?
2 replies →
Astra is 58%. The current title says it's "rivaling Astra"
1 reply →
DeepSeek v4.1 Flash 31.2%
Source: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash#compa...
1 reply →
This explains a lot about Sonnet 5.
So, better than Sonnet and Luna? lol
This is why I find benchmarks absolutely worthless.
First, almost all models are within spitting distances of eachother.
Second, it never translates to being better for my own workloads.
You just need to make your own benchmarks.
For comparison Qwen 3.8-Flash-Next which runs in under 190GB of RAM locally scores 25.3% on terminalbench 4.0.
Yeah DeepSeek V4.1 beats this by 15 percent (4 points) on terminal bench 4.0
When a benchmark becomes a target, it's no longer a good benchmark...
people are just fighting for numbers.. i don't fundamentally see the model being better
Came here to say the exact same thing! People have to stop paying any attention to coding benchmarks that aren't Terminal Bench 4.
I noticed I noticed they didn't include Gemini 3.8, which also murders DeepSWE and Terminal Bench 2.0 -- because they are useless benchmarks now!
Of course in a couple months TB4 will also be old hat, so TB5 will have to be the new real benchmark.
Your post made me wonder if Artificial Analysis had finally moved to TB4 and lo and behold they have and Astra is tied with Fable 5.1 at 53.
That then made me realize that they lower the bars of tied scores so on the site it looks like Astra in second place. Weird. Anyway, yes, so many of these composite benchmark sites are irrelevant if they're not trimming the fat and sticking to the most up-to-date variants.
Yeah, this echoes my thoughts. I will be very surprised if a model with 2.8T parameters reaches the intelligence and capabilities of 10T parameter models. RL can take things far, but not that far.
Closed weights AND benchmaxxed. Somehow this company raised 2bil at a 48bil valuation. Pure insanity. I feel bad for their investors (not really, but... Still). Andreessen Horowitz is being played like a fiddle.
4 replies →
Cognition, the same company that a few years ago demoed a coding bot purporting to be able to autonomously complete upwork tasks, but upon closer inspection was going off the rails and not even completing what was asked?
https://www.youtube.com/watch?v=tNmgmwEtoWE
As others have mentioned this is post trained from Kimi k3, which is already quite capable, so it can't be that bad, but any claimed improvements in performance should be taken with a grain of salt.
A friend recently pushed me to try out their coding platform, Devin, after I decided to move away from Cursor. I had the same reaction: "What, the con artists from like 2024?" But after some cajoling, I gave it a shot and was pleasantly surprised. I guess they learned their lessons, grew up, and are doing good work now, maybe?
Their ads in SF are pretty funny. “Remember Devin? It’s good now”. Okay, very self-aware, Cognition.
Devin CLI is easily the worst-in-class coding harness I've ever been subjected to using, rife with bugs (up to and including dropping answers to the question tool), so I can't say I'm exactly inspired to try anything else the company produces.
2 replies →
Not all that surprising. The original Devin really was just an early attempt at agentic coding before models were really even trained for it. Now that it's a well established pattern and we've figured out what works, I'm not surprised they've morphed into something reasonble.
When they launched Devin it was supposedly at the performance of an engineering intern. Friends who used it found the bad parts of an intern (tons of handholding, review required) but it didn’t learn from mistakes or add throughput.
They seem to love a good overpromise.
I think the evolution of the harness and ability to preserve loop context outside the context window has made running these kinds of agentic experiences easier.
sorry so many buzzwords to say, the capabilities to do this kind of work are more accessible and easier to manage, so now it works!
Good to see, and agree they were severely overhyping their product back then.
Devin lacks many features and doesn't even have an exec mode. Other than they subsidize the subscription, the added value here is low.
2 replies →
a reminder that money does enable you to make mistakes and buys you the ability to correct from them
1 reply →
As others have mentioned, it's really matured a lot and at this point is one of the best cloud-hosted, team-managed coding agents, when factoring overall UX, testing and QA lifecycle via its sandboxes, and its ability to be controlled with an API. We use it quite heavily.
It's coming from a different starting place than Claude Code or Codex are as individually controlled single-developer tools. Devin has been more persistent in pursuing the direction of something that operates more autonomously at the team level, as a peer. And while it might be slightly behind in raw harness ability (maybe?) it's probably ahead on the team-focus.
Your startup is based around AI coworkers, so I expect you are biased towards overvaluing their usefulness.
1 reply →
I'm surprised they can post-train a closed model off of K3. Is that the norm for open-weight licenses?
Yes. If they are really open, you can use them as you wish. Some licenses are less open though.
1 reply →
Well the models did get better but yeah their early product was godawful
Previous versions were based on Kimi too. I'd consider if it I could access the model outside Devin. No lock in for me, thank you very much.
Where are the model stats? Is this open-weights? If not, why would I use this over DeepSeek Flash 4.1?
I think these competing labs need to realize that no one wants another closed-weight model provider... We aren't even happy with the two we have right now, and their days are entirely numbered. If DeepSeek 4.1 flash is really as good as it's benching, we're probably a month away from 1/3rd of users moving off the closed-weight models in favor of something they have more control over (or is cheaper).
The big labs love to release their new model and quantize after the first week. You don't have that problem using dirt cheap API rates on OpenRouter. DS 4.1 flash is also faster than fast mode Astra. OAI's subscription rates are good value, but now these new open-weight models are nearly as cheap on API usage rates. I honestly can't wait for the day we're not beholden to the two big labs anymore. No wonder there's so much fear pumping happening at the moment from Anthropic and their funded NGOs.
This really just exists so cognition can stop spending API tokens with Anthropic or OpenAI.
Basically any successful AI based service will do this because at scale the frontier models are expensive and you’ll have enough data to fine tune your own.
Same reason Harvey is doing models now and basically every other provider
Couldn't they just grab and run an open weight model to save on API tokens?
1 reply →
> If DeepSeek 4.1 flash is really as good as it's benching, we're probably a month away from 1/3rd of users moving off the closed-weight models in favor of something they have more control over (or is cheaper).
Only in the world where the incumbents don't react. Eg if they saw lots of users moving away, they'd drop prices or do something else.
1M cached tokens on deepseek is $0.006, the big labs can't sell anywhere close to this, they have funders expecting returns and huge overhead.
btw I've had a ton of fun with the new deepseek today, I was waiting for my OpenAI 5h limit reset and decided to give it some problems for fun, got pretty great results. Tried some harder problems and still got great results. I don't expect it to be Sol class or anything but I really didn't expect it to be anywhere near this good so we'll see where it ends up. And it's really fun throwing crazy amount of tokens at the wall for ~free instead of watching the subscription limits tick closer while your agents churn away.
5 replies →
OpenAI just paused new subscriptions to their $200 plan. They are in a rock and a hard place. Obviously the Astras and Fables of the world are exponentially more expensive, but for...less than exponential returns. The question is whether they can leverage the marginal advantage into something that justifies the diminishing returns before the bottom catches up to them.
On the one hand you, if you bought a lot of compute a couple years ago (perceived demand, perceived shortage) you are in a good spot temporarily. But the counter to that is that everyone else is becoming more compute efficient so maybe that advantage isn't what people thought it would be. I can almost, almost run DS4.1 Flash at home. 4 sparks can do it at 200+ tokens per second. I have two Sparks, so I am not in the club. Neither is your average laptop owner or gamer either. But your average HN software engineer can probably easily swing 2 sparks.
4 replies →
> we're probably a month away from 1/3rd of users moving off the closed-weight models in favor of something they have more control over (or is cheaper).
DS 4 Flash requires large amounts of memory to run at reasonable quants (I think a system with 160 GB or so). DS 4.1 Flash is even larger, I think around 250 GB.
Any DS version is dumb when compared (in realworld tasks) to Astra/Opus 5, which means, one would spend thousands of dollars, and still need to rely on cloud services to do jobs that are non trivial.
Not sure it matters when devin is the most consistently shit product I've used. And yes, I tried again, they wasted the money on the billboards.
> And yes, I tried again
This made me laugh a bit. I was forced to do an evaluation of their shit product twice due to being backed by the same PE firm; "take a look at it again, it's much better now". It sucked the second time also...
I have not met a single person/company that uses Devin… does anyone here actually use it?
Unfortunately yes. It's awful, other than that they at least support both open-weight and proprietary models to route to. So I guess their model router is "fine", but the harness? As I said in a sibling thread on this page:
> Devin CLI is easily the worst-in-class coding harness I've ever been subjected to using, rife with bugs (up to and including dropping answers to the question tool), so I can't say I'm exactly inspired to try anything else the company produces.
I pay for it (mostly because they grandfathered me from the old prices)!
I used to use windsurf as my main editor until they changed their pricing model. Now i use it just to burn my weekly tokens on fable/astra if i remember to that on a task and that's it.
I only used their "DeepWiki" automatic docs, they are pretty decent at getting an overview of a large project and are relatively accurate, with diagrams and anything. Haven't tried out their coding agent stuff.
My company uses it. We have a bunch of seats in an enterprise plan and have been using it for 9 months or so.
It's a great product compared to Copilot. It is also the first AI tool I used heavily outside of creating random images or one off questions.
I'm now using all three, Devin, Claude, Codex. I'm finding Claude and Codex to be much better. One of my biggest gripes is that the web client and desktop client for Devin are two completely different harnesses, so the quality of responses varies greatly.
I use it, and have been happy with it for the most part. Like sibling, I use it for GPT-5.6-Sol and Opus work, and use their free models (GLM 5.2 for the past few months, trying SWE-2 now).
my friends at infosys are being trained on it
1 reply →
Well, where VC money is involved, the aim of these ads placed in SF is not exactly to convince any potential users.
> SWE-2 is post-trained from Kimi K3
On the one hand I would have expected a completely new model, on the other hand it's an RL-ed K3 go Fable 5 capabilities, which demonstrate that this is probably possible, which is nice.
why are all American AI models basically Kimi in a trench coat
No one in the US is going to fund pretty good open source with VC money.
US has OpenAI/ Anthropic/ Google/ meta/ SpaceX atleast trying to make frontier foundation models 2 are using VC money + cash flow, last 3 are mainly cash flow + equity and debt.
China has state banks and similar willing to fund lower margin open source labs.
It's (one of) the best freely available?
Seems like benchmaxing? For example for Terminal-Bench 4 it doesn't have great results. And why not show other benchmarks?
Probably, FrontierCode is made by Cognition itself. The model also seems worse in every way than DeepSeek v4.1 Flash, launched today.
Also the submitter's account is very new which makes me suspicious of self-promotion.
SWE 1.6 was great for small tasks. Very fast and good enough. 1.7 was unusable for me. Took more time thinking than GLM 5.2 and seemed to be generally running in circles. I tried it but abandoned it.
Looking forward to 2 -- maybe it'll be usable
I am skeptical. Lived experience is what matters and I don’t have anyone in my life (Devin shop) saying good things about SWE other than it’s free. Hope I’m wrong and it’s not so bad this time
I think I'm probably in the minority here, but for my line of work, the software engineering and coding is only a small part of the work. I write simulation software, so a deep understanding of physics, math, and how they can be applied to the software is absolutely crucial. I'm assuming this model is tuned to be more focused on SWE topics, and the very reason we seek "multidisciplinary" hires is the also why I actually need a jack-of-all-trades model to back my coding agents.
I'm also in simulation software! Wondering which models you are finding helpful, the models I'm using for general SWE skills are horrible at our simulations and even basic physics/engineering calculation and intuition
I use Claude Opus 4.8 almost exclusively. I have had fairly good experiences with it. One time it derived an entirely novel simulation method different than anything in literature by combining its knowledge about how problems in other fields with similar underlying mathematical structure are solved. That was a bit of a Jacobian Conjecture moment for me.
I'm in the same field and I find GPT to be better at understanding physics conceptually but Claude is better at writing numerical code. I use cursor so many of my sessions start in GPT and switch to Claude.
Post made by account 2 days ago.
Is there an implication here I'm missing?
I like Cognition as a company and hope they succeed. Seemingly excellent engineering org.
I used to really like Windsurf. (Now Devin. Kind of? But also now Antigravity.) I still use it as my editor but haven't touched the agent for a while simply due to the rise of Codex.
Their benchmark used to show other metrics, like output tokens and time, but now only shows cost:
https://cognition.com/frontiercode
Which is too bad, since all of the gains here appear to be from massively reduced output tokens?
The model SWE-2 is based on, Kimi K3, is cheaper per token than Sol, but costs more per task (ArtificialAnalysis) due to using way more tokens.
Whereas, based on the graphs, SWE-2 appears even more token-efficient than Sol! That might have been worth showing off, if true.
Wonder if this was the model that drove factoring the rsa-260
The write-up from yesterday was by somebody from cognition using Devin to translate existing cpu sieving methods to gpu and to optimize the gpu sieve.
Unless it can do CAD via computer-use how can you say it rivals GPT-Astra?
If everything basically rivals Fable, then why is everything still using it for comparison?
Have you spent at least 10 seconds thinking about it or are you asking just out of spite?
Let me grab my calculator and add up the time I have spent reading about model releases since Fable has been released. It seems they all place themselves relative to Fable. I'm sure that time has added up to far greater than 10 seconds. At some point, it ceased to be a meaningful differentiation. This is especially true when I put the model through real usage.
Why doesn't clickbait trash like this get moderated ?
Because it's not clickbait trash / doesn't violate the site rules?
Dang is pretty good at enforcing stuff. There's a flag button and you can email reports if you're really bothered.
Fair enough that they did a "propoganda and censorship" eval but not sure why i'd care about that in my highly juiced SWE kimi FT.
Please correct me if I'm wrong, but this appears to require Devin to use? I'm disappointed to see I need to use a bespoke platform to interact with this agent, to the point that I probably won't be trying it.
SWE-2 is free to use for users like yourself for the next month, and almost all usage should be supported via our CLI (https://docs.devin.ai/cli)
:)
Disclaimer: I work at Cognition, although was not involved in SWE-2
But I don't want to use your CLI. I already have my own harnesses and workflows. The friction is too high to "just try out" a new model like this. It would be preferable if I can evaluate it over, say, open router like all the other models and then decide from there if it's worth downloading a bespoke tool chain for only 1 lab's models
1 reply →
Your own CLI? Not even a /v1/chat/completions API? Is your business model based on pretending LLMs are not an interchangeable commodity already?
1 reply →
I just gave it a try and it doesn't appear to be free, it used up some of my on demand usage. It does say 75% off though. Seems like for Pro subscribers SWE-1.7 is free, maybe SWE-2 is free for them?
1 reply →
Hey man, how’s the Poke SOC2 audit going? Must be any day now that it’ll be finished, right?
Heads up: it doesn't appear to be available on the Devin CLI (for me as a free user).
If its weights are open, that covers a multitude of other sins. Sufficiently-strong performance on the part of the new model would justify adapting existing tools to work with it.
Incest in human biology causes mutations that's bad in the long term.
Same for AI models trained on Kimi-3 or other models like Chinese models do. They suffer from the same issue.
As an Econ graduate, pretty cool seeing Pareto in the "AI-bro" zeitgeist. Slightly surreal watching a 1906 welfare economics idea get rediscovered as a plotting convention. The original, if anyone fancies 579 pages of Italian: https://archive.org/details/manualedieconomi00pareuoft. There is an English translation somewhere.
Pareto frontiers are pretty commonly invoked to describe tradeoffs in computer science and have been for quite a while. I remember the term being used in one of my early algorithms courses to describe the tradeoff between data structures with fast writes, ones with fast reads and ones that tried to balance the two.
IIRC cognition boasted about hiring a lot of competitive programmers and algorithms experts back when they released Devin, so it tracks that they'd use the term.
It's interesting watching people throw about pareto frontiers sort of like how RF nerds approach the shannon limit (in a practical real world sense of the term, like charting possible modulations/data rates on a two way satellite modem's manufacturer datasheet).
Pareto leaked out of the sociology/econ bubble a long time ago :) Pareto principle, Pareto efficiency, Pareto distribution have been in the pop-sci buzzwords for quite awhile, I probably encountered it first in the 4-Hour Workweek. I don't think you can read a self-help book without the author introducing it as a groundbreaking principle to live your life by.
"SWE-2 is post-trained from Kimi K3"
Yeah, I'd expect model performance to be super spiky on SWE work, at least they admit it with the name of the model. It's distilled from an already-distilled model.
Maybe still worth it if their "64% cheaper" figure holds.
I don't think you know what distill means
4 replies →
With the Devin subscription even at the 20$ plan, they offered unlimited SWE 1.7 usage. Wondering if they do the same for SWE 2.
2 replies →
I presume post training is significantly easier than the distillation/training the top Chinese labs are doing.
I wonder if, similar to the American labs, they'll become stingy with their weights once they start getting immediately undercut by a wave of slightly better derived models.
At work I setup a cloud worker, where i can spin up as many concurrent agents I want, with unlimited fable 5.1 (thanks employer!!).
I now just work from my phone, and speak into the agents as they run. I dont write code and I dont write documents. I work on very complicated distributed systems. I dont open my laptop most days. Its a legacy brick I carry around.
Some of my coworkers are still doing things by hand, and are working long hours to produce 25% of the output (when considering hours worked). I stay quiet with my setup. We are in the end times for this job for the people that can see clearly how to automate their own job
that's why after 1 year of product development of these AI 20x maxxed speed, we reached AGI 'wizards', there's really no difference in output, outstanding bugs no longer get solved and sites still suck, even doing things that were just regular development 20 years ago. Are you sure they aren't only producing 2.5% of your output that you manage just by farting into your phone? Are you sure it's 25% really? Seems way to high, days when I have diarrhoea my AI agents move even faster
please keep thinking this so i can relax with my automated job
4 replies →
[flagged]
Not avail on open router?
Do they have anything as good as Amp (which is able to use free models)? Amp has been in the lead for almost a year now and doesn't seem to be relinquishing it.
Well written and good diagrams. No idea the verity of the TMBB (trust me bro benchmarks) but it was pleasing to look at
I'm using Codex, Gemini etc, they all have desktop apps and have a plan, how do i use SWE-2? Thats is a problem they have. I'm not about to switch out my workflow and plans with a shiny LLM that looks benchmaxxed and graph maxxed.
SWE-1.5 was surprisingly good when I used it last. I feel like Cognition is one of the solid players that’s flying a bit under the radar while Anthropic and OpenAI race to IPO.
Odd to get downvotes simply for sharing my experience. Like it or not, Cognition has a good frontier model and they are building serious products. They are working hard which is how you become successful. Sorry if that ruffles your feathers.
The horrible website is made by Claude or Cognition is distilled. I'm so tired of it all.
all this while Cognition doesn't even make the top page of search results, that's some impressive stealth.[1]
(just to be clear, I am a far-left activist who spends most of my time working on funding Social Security Trust Funds (OASI & DI Solvency), which could impact my search results - this was while I was logged in.)
[1]https://www.google.com/search?q=what%27s+cognition+in+ai or https://imgur.com/a/UdxtnGg
[dead]