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.
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 guess I don't. Does post-training from another (larger) model not fall under the umbrella of distillation? I'd imagine it leads to the same spiky-ness issues...?
Distilling you don't have the actual model weights of the teacher. All you have are the teachers answers to a lot of questions. You then teach your own smaller model to answer more similarly to the big teacher model.
Fine tuning you have the actual model weights of the original model, you then train that model to answer in a different (or better) way.
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 derived models.
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
It’s preferable to keep inference capacity available for users using main Devin products than openrouter atm. Might change in the future. Even OAI is cutting off new plan signups to keep up with demand.
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 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.
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).
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?"
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.
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.
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?
> Altman was caught in previous attempts trying to game benchmarks
Sounds like something you just made up, or maybe you read it on some other Reddit/HN post and started repeating it because it aligned with your biases.
> does anyone believe that he's found his moral compass and decided to stop exploiting as much as he can get away with?
I don't think "OpenAI" is equivalent to "Sam Altman." I think if OpenAI was intentionally "benchmaxxing" purely for marketing purposes that information would leak, because OpenAI is full of good-faith researchers (although it can be difficult to avoid overfitting even if you're actually trying to improve the model's general abilities)
And lastly I think anyone can actually try Sol themselves and see that's it a good model, or if that's too subjective, it is clearly better than the previous version. The benchmarks are reflecting actual progress and anyone can verify this themselves.
There's this whole discussion going on about agents being more independent now. They don't follow instructions so well, they continue until the problem is done (sometimes too long), they don't ask the user for feedback.
That is a kind of benchmaxing: they are made to complete benchmarks tasks and one-offs well, and no longer work well in tandem with the user.
Regardless what you call it, it's a divergence between what the power user wants and what the model developers want, I think.
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.
> 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, 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 the top rope with an emphatic 92.4%. this is the definition of bench maxxing.
> 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?
Wait a second, are we taking into account the massive difference in terms of resources of these two companies?
You should use my model then, I spent about 30$ in electricity and used my existing RTX4090. It is not very good, but can you compare it with others really? You can use this service via a private API with a VPN, email me your credit card details for access.
Wouldn't work. A dynamic but verifiable problem. Is just a perfect target for a RL environment. If you don't have the verifiable part the benchmark is useless, or really expensive with human review. (Or just open ended)
There doesn't need to be a single correct response. Generate "novel" problems with a set of acceptable solutions/outcomes, then verify that the response satisfies the criteria.
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.
You can't justify it being OK just because it's common. Here, this just makes benchmarks into a low signal and useless marketing number once people get numb to all the 99%s.
Also no idea what viruses and cancers have to do with this. Cancer is surely very poor reproductively because they never spread to other hosts.
Cancer is simply cells breaking free of the cooperative jail. Essentially the grey goo scenario of nano machines. Instead of cooperation they just do their own thing.
Cells have a certain optimised DNA mutation rate kept in check by various machinery. Multi cellular life expects each of these little replication machine to co-operate in the grand scheme of running a body. But it's also required in the grand scheme for DNA to mutate a little bit to ensure population variance. So you could say that cancer is the tax paid for having cooperative yet flexible and adaptive nano machinery.
So yes the propensity for cancer developed under evolutioniary pressure towards a non zero level.
The population could have optimised for zero cancer but it would not have paid for itself in terms of overall population adaptability and survival.
> Also no idea what viruses and cancers have to do with this. Cancer is surely very poor reproductively because they never spread to other hosts
The host's survival is not the benchmark of the reproductive unit, which are the cancer's cells short-term reproduction.
The distinction is the reason benchmaxxing is in fact not good: the benchmark is only approximately related to what people actually care about. You do want your cells to reproduce sucessfully, after all; you just also want some emergency stop buttons for when they go wrong, and those things failing is your body's benchmark rather than your cell's benchmark.
(There's at least two examples of cancers that can be spread from host to host; lupine genital and taxmanian devil nasal, IIRC)
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?
It is rivaling Astra, on their own benchmark that they made (FrontierCode), that they ran themselves in their own closed-source ecosystem that isn’t reproducible by anyone.
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?
Terminal Bench 4.0 did not introduce new questions. All tasks were public for a while. If you look at GLM 5.2, which is using the same base model as 5.3, but was released prior to most tasks, it does extremely horribly on terminal Bench 3.0 (4 to 8 times worse than every other model) - source: https://benchlm.ai/benchmarks/terminal-bench-3
Opus 5 generates really good code and terminal commands though. It's just bad at the accompanying text it tells you. These benchmarks don't grade the text generation of the response I don't think, only the task outcome.
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.
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.
I'd be pretty surprised if someone told me few years ago communist China, state banks would become the main founders of open source compute and our last hope against monopolists like Musk, the whole bunch at OpenAI and so on.
To be fair Zuck is releasing fairly capable open source models, but Chinese labs are way ahead.
> 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.
I did a lot of that kind of work with the older version of DeepSeek before they upped the prices.
For example, it was quite good to get a decent Sashiko review. Sashiko is a Linux kernel review agent with interchangeable LLM driver. It's very good, but it eats tokens like crazy.
> 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.
Losing most of your customers tends to sharpen the mind a bit. They could eg stop pushing out the absolute frontier for a while and focus on making what they have run cheaper. Or they go and do more lobbying against China. Or a million other little things that take more than 30 seconds to come up with when writing a HN comment, but less than a week for someone who's smart and paid to do this for a living.
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.
It's not particularly good value if you are just comparing $ with no other context. Point is just that it's accessible and so now Anthropic and OpenAI need to make both a performance proposition and a value proposition.
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
> 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 versin 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 not trivial.
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
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 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.
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 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).
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 not 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.
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.
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...
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?
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.
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.
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.
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.
Had the same experience on it when it was going by Windsurf. Had to wire up a skill hooked to terminal runs or else it would hang or freeze and never finish literally every single time
I see issues with other harnesses too but not with the regularity I was getting from this. And the one moat they had with the better UI for per-project multi-agent orch disappeared and now is standardized
Can't say I've tried the CLI. I've mostly focused on the cloud agents, which I was explicitly looking for. I compared against cursor's cloud agents, ampcode, and hoplite, and came out surprisingly enjoying devin.
I will say that the lack of parity between Devin cloud and Devin desktop is downright embarrassing. It's very clear that the latter is a thinly-reskinned Windsurf. A visually similar UI with vastly different capabilities. Definitely a black mark on the whole thing.
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.
Yes, perhaps fair. But my point was somewhat narrow. I wasn't saying that team-managed agents are good to go for all cases and that they're better than individual dev-managed ones. Just that they have gotten better and that of those Devin has some of the better UX.
Our experience might also not be typical because we have built infrastructure around making Devin and similar agents work better. But no financial ties to Cognition or anything. Just pay them too much as a customer.
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.
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.
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.)
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
173 comments
[ 1.8 ms ] story [ 14.1 ms ] threadMaybe still worth it if their "64% cheaper" figure holds.
Fine tuning you have the actual model weights of the original model, you then train that model to answer in a different (or better) way.
What you're describing is just synthetic data.
Note Anthropic misused the term in their post about Chinese model distillation, deliberately I assume.
https://arxiv.org/abs/2106.03310
https://arxiv.org/abs/2207.12106
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 derived models.
:)
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.
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?"
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.
Sounds like something you just made up, or maybe you read it on some other Reddit/HN post and started repeating it because it aligned with your biases.
> does anyone believe that he's found his moral compass and decided to stop exploiting as much as he can get away with?
I don't think "OpenAI" is equivalent to "Sam Altman." I think if OpenAI was intentionally "benchmaxxing" purely for marketing purposes that information would leak, because OpenAI is full of good-faith researchers (although it can be difficult to avoid overfitting even if you're actually trying to improve the model's general abilities)
And lastly I think anyone can actually try Sol themselves and see that's it a good model, or if that's too subjective, it is clearly better than the previous version. The benchmarks are reflecting actual progress and anyone can verify this themselves.
That is a kind of benchmaxing: they are made to complete benchmarks tasks and one-offs well, and no longer work well in tandem with the user.
Regardless what you call it, it's a divergence between what the power user wants and what the model developers want, I think.
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.
Yes? Just like every single model from every single AI lab.
A model that was released a couple months ago scores 50% higher than SWE-2, 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 the top rope with an emphatic 92.4%. this is the definition of bench maxxing.
Wait a second, are we taking into account the massive difference in terms of resources of these two companies?
You're either competitive or not.
Yes.
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.
Yes! extremely sharp RL-fried model. byte perfect hash gates and soak and smoke tests abound.
It's not groundless, but it's not entirely accurate.
You can somewhat control for the change in benchmark difficulty by comparing changes in score from model A to models B, C, D, etc.
For example, the benchmark scores for GLM models dropped much more significantly than ChatGPT models when transitioning from TB2.1 to TB4.
IMHO, evidence like that is the clearest sign of "benchmaxxing" that we can point to.
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.
Also no idea what viruses and cancers have to do with this. Cancer is surely very poor reproductively because they never spread to other hosts.
Cells have a certain optimised DNA mutation rate kept in check by various machinery. Multi cellular life expects each of these little replication machine to co-operate in the grand scheme of running a body. But it's also required in the grand scheme for DNA to mutate a little bit to ensure population variance. So you could say that cancer is the tax paid for having cooperative yet flexible and adaptive nano machinery.
So yes the propensity for cancer developed under evolutioniary pressure towards a non zero level.
The population could have optimised for zero cancer but it would not have paid for itself in terms of overall population adaptability and survival.
The host's survival is not the benchmark of the reproductive unit, which are the cancer's cells short-term reproduction.
The distinction is the reason benchmaxxing is in fact not good: the benchmark is only approximately related to what people actually care about. You do want your cells to reproduce sucessfully, after all; you just also want some emergency stop buttons for when they go wrong, and those things failing is your body's benchmark rather than your cell's benchmark.
(There's at least two examples of cancers that can be spread from host to host; lupine genital and taxmanian devil nasal, IIRC)
The truth is much more mundane. It's just Goodhart's Law.
Some might say their job is optimizing ways to make numbers appear better without substantial change in input or output.
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?
They are all gaming these benchmarks, it is perfectly reasonable not to trust any of them.
Probably
- Sonnet 5 - 12.4%
- Luna - 17.3%
- Grok 4.6 - 20.3%
- Sol - 37.3%
- GLM 5.3 - 41.8%
- Opus 5 - 51.8%
Source: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash#compa...
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?
Seriously.
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.
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.
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.
An aside: When did talking like incels became cool?
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.
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.
To be fair Zuck is releasing fairly capable open source models, but Chinese labs are way ahead.
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.
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.
I'm willing to tolerate babysitting things a lot more if I know I'll get almost instant results.
But great that we have a new leader in performance/price in that segment.
For example, it was quite good to get a decent Sashiko review. Sashiko is a Linux kernel review agent with interchangeable LLM driver. It's very good, but it eats tokens like crazy.
Losing most of your customers tends to sharpen the mind a bit. They could eg stop pushing out the absolute frontier for a while and focus on making what they have run cheaper. Or they go and do more lobbying against China. Or a million other little things that take more than 30 seconds to come up with when writing a HN comment, but less than a week for someone who's smart and paid to do this for a living.
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.
That's like four years of ChatGPT + Claude subscription.
Eight years if only ChatGPT, or sixteen years of the Pro 5x subscription.
I'm not sure? If we have techniques to use the hardware even better, that will make the hardware even more valuable, won't it?
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 that were the case they'd just be running Qwen or GLM on their own metal. This exists to "justify" their valuation.
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 versin 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 not trivial.
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
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.
Looking forward to 2 -- maybe it'll be usable
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.
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.
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 not 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.
> 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.
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...
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.
Or at the very least, make more mistakes.
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.
They seem to love a good overpromise.
I see issues with other harnesses too but not with the regularity I was getting from this. And the one moat they had with the better UI for per-project multi-agent orch disappeared and now is standardized
I will say that the lack of parity between Devin cloud and Devin desktop is downright embarrassing. It's very clear that the latter is a thinly-reskinned Windsurf. A visually similar UI with vastly different capabilities. Definitely a black mark on the whole thing.
Everything feels dull and they're always several features behind while Cursor is just killing it every other week.
I'm back to VsCode plus Copilot Pro.
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.
Our experience might also not be typical because we have built infrastructure around making Devin and similar agents work better. But no financial ties to Cognition or anything. Just pay them too much as a customer.
Dang is pretty good at enforcing stuff. There's a flag button and you can email reports if you're really bothered.
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.
(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