Overall an improvement over Opus 4.8, but I'd still say Gemini 3.1 Pro has more of an artistic vision even tho it fails tool calls and writes buggy code sometimes.
Ik almost everyone is interested just in the SWE stuff, but this has been a good eval for me to think about how big the model is, how "creative" it is for generating new ideas etc.
For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Most of the money right now is in coding. Openai and Anthropic just have to be 6 months ahead of SOTA open source models and they'll capture most of the enterprise and dev market
Yes I'm an engineer (20 years most in games/graphics industry) and only use it for code. I've been using glm 5.1 this week a lot. I went in expecting another "decent" but not really "up to standard" open source model.
I highly doubt I'll ever use Claude again.
I think you are wrong about Claude being any significant level better
I've been mostly coding with GLM-5.1 as well and I agree with you. DeepSeek V4 Flash is another very good surprise. Incredibly cheap, fast and effective.
Well I think there are a multitude of harder measurements that would disagree with you, but ultimately there is absolutely a use case for cheaper open models (or even cheaper tiers of proprietary models) and in fact the unsolved optimization everyone is trying to get to is how much spend to use for a given task. But there will always be a market, especially in enterprise, for the best performance there is to offer
I strongly disagree. I'm an engineer - I'm all about the fastest, cheapest thing that meets the requirements. I don't need Opus 4.7, even for my complex programming tasks. It costs over 10x other models available that still give good enough answers. Those smaller models are also a lot faster to output tokens, which saves me time.
Once the model gets good enough, the returns on bigger models diminishes quickly. I don't want to spend 10x the money and wait 5x the time to get answers that are equivalent.
If I generate code with Claude, ChatGPT, and GLM 5.1, I can't say which model is which reliably. I exclusively use Claude more out of superstition than reason.
> For coding you always want to go with the best model in the category
This is transparently false, because the best "model" is still competent human developers. They're just more expensive. If you're willing to use current LLMs at all, it means you're willing to sacrifice quality for a better price, and your disagreement with the comment you were replying to is entirely about what the optimum tradeoff is.
Maybe, but that's a different claim than the one I was responding to. And also raises the question of "if the lower quality but cheaper output of frontier models is more cost effective than humans, is the even lower quality but even cheaper output of OSS models is more cost effective still?" With an absolute rule like GP suggested ("no, you always want the best code generator") the answer is clear, but it get much murkier if you reject such rules (as you have to to be an LLM coding proponent)
It was true 6 months ago, not anymore. Frontier models now outperform developers on many tasks, be it on quality/readability/maintainability, and let’s not talk about speed…
I've seen the code they produce without extensive help from human developers, this is clearly false.
Good to see the classic "yeah the models weren't good enough six months ago, but this time they actually are, promise! Please forget you were hearing the exact same thing six months ago!" is alive and well though.
Are you aware of performance trends though? You’re painting a picture that seems to ignore how things have consistently trended for many years now, even pre ChatGPT. It is absolutely data driven to say “an inflection point has happened within the last 6 months”. And that was also true 6 months ago (where people started using coding agents fairly consistently since sonnet 4). And it was true 6 months before that. It’s not like people are like “we’ve fixed all the bugs!” And then nothing has changed. I don’t necessarily agree with the parent poster that agents are better than humans but they are certainly much better at many tasks.
> Are you aware of performance trends though? You’re painting a picture that seems to ignore how things have consistently trended for many years now, even pre ChatGPT.
Models have been getting better, but all that follows from that is that newer models tend to be better than older ones. It doesn't follow that they have (or even will in the future) gotten better than anything else, be that human developers, a given definition of good enough, etc.
> It is absolutely data driven to say “an inflection point has happened within the last 6 months”.
With all due respect to OP (who I think is responsible for popularizing that way of phrasing it), I don't think it is when you consider the actual definition of "inflection point". At best I think you can say that models crossed a lot of developers definition of good enough around then, which is a different thing. The problem I have with that is that as a (mostly) outsider looking in, it doesn't seem like they're right.
> Models have been getting better, but all that follows from that is that newer models tend to be better than older ones. It doesn't follow that they have (or even will in the future) gotten better than anything else, be that human developers, a given definition of good enough, etc.
But this is not true, you’re saying we only have relative performance numbers and not absolute measures of capabilities and reliability but that’s simply not true. OSS benchmarks as well as the internal flywheels of these companies are good complementary measurements.
> At best I think you can say that models crossed a lot of developers definition of good enough around then, which is a different thing
That’s the inflection point. Implication is a massive jump in adoption. We’re not like pulling this out of a hat, there are a number of compelling datapoints. The onus is on people to bring actual evidence that contradicts all of the data and observations we have.
> you’re saying we only have relative performance numbers and not absolute measures of capabilities and reliability but that’s simply not true.
No, I'm saying that the claim you were making ("current models are better than some non-model based standard X") does not follow from your premise ("current models are better than past models"). It's possible that your claim is still true (although I don't think it is for most of the values of X that matter), but that wouldn't change the fact that the argument made is invalid.
As stated, your argument was basically the classic "my 3-month-old is now twice the size he was when he was born" meme, except if the tweet claimed that the kid currently out weighed an elephant.
> That’s the inflection point.
No, it isn't. An inflection point is when the direction of curvature changes. If we crossed over into the diminishing returns part of the logistic function, that would be an inflection point (as would the case where we had been in the diminishing returns regime, but then progress went back to speeding up).
> Implication is a massive jump in adoption.
The point I made was that "a massive jump in adoption" doesn't actually imply "the models are actually good enough now", only that a lot more people think they are.
OK I am having the wrong conversation, that you are right -- parent OP saying
- best model is still a human: this I SORT of agree with, but like I say its uneven
- response is: "this was true 6 months ago but is now false" -- that is sort of a mixed bag; if its saying we can now replace SWEs thats demonstrably wrong, if its saying it now clearly has superior abilities in many parts of the SWE workflow then its demonstrably right. I would argue this has been true for longer than 6 months.
- you say: "I've seen the code they produce without extensive help from human developers, this is clearly false." -- I agree with you that you need to help coding agents substantially, but I think at this point the convo is unclear what anyone is actually addressing or responding to
> No, I'm saying that the claim you were making ("current models are better than some non-model based standard X") does not follow from your premise ("current models are better than past models").
that isn't my premise though, but I admit I misread you. I think you are saying "people keep saying it's good enough to replace SWEs(?)" every six months and they are wrong every time. I don't disagree that we have not gotten to a "we dont need SWEs anymore" point, but I think its a bit of a strawman: who is making the claim you are addressing?
> As stated, your argument was basically the classic "my 3-month-old is now twice the size he was when he was born" meme, except if the tweet claimed that the kid currently out weighed an elephant.
No no, I'm talking about performance in absolute terms. These are strong proxies (SWE-bench, etc) though they have serious limtiations. The "yea but when is failure rate low enough for us to replace an entire tranche of processes" thats a harder question to answer but the strong proxy for that is adoption.
> The point I made was that "a massive jump in adoption" doesn't actually imply "the models are actually good enough now", only that a lot more people think they are.
No but then the point I'm making is we're drifting further and further away from Occam's razor.
> No, it isn't. An inflection point is when the direction of curvature changes. If we crossed over into the diminishing returns part of the logistic function, that would be an inflection point (as would the case where we had been in the diminishing returns regime, but then progress went back to speeding up).
I admit inflection point may be the wrong term here, but I hope you know at least what I'm trying to say; maybe like regime change or something. But plenty of data supports a major change around Nov to ~Jan. revenue, weekly active users, business subscriptions, GitHub commit estimates, you pick which is your favorite data source but they all are complimentary and all point to the same thing.
First, to clarify my own position here: I use LLMs for code review, to help with some planning, for the occasional throwaway prototype, and as a more advanced rubber duck, but I do not let LLMs write code I care about, even with human review (because human review is imperfect).
> its uneven... if its saying it now clearly has superior abilities in many parts of the SWE workflow then its demonstrably right.
In what ways are current LLMs better excluding speed and cost (because getting those things by relaxing constraints on quality has always been trivially possible)? Even the fabled (heh) Mythos seems to be at best roughly equivalent to a competent human security researcher.
> I think you are saying "people keep saying it's good enough to replace SWEs(?)" every six months and they are wrong every time. I don't disagree that we have not gotten to a "we dont need SWEs anymore" point, but I think its a bit of a strawman: who is making the claim you are addressing?
Most of them aren't saying that the models are good enough to full replace developers, but this definitely isn't a strawman. I've been seeing the same basic claim for at least 18 months at this point.
> No no, I'm talking about performance in absolute terms. These are strong proxies (SWE-bench, etc)
Unless you have some non-LLM scores to compare to, those are still relative measures. They show/suggest that LLMs are getting better (at least in some ways), but without a definition of "good enough" in the same metric, that isn't sufficient to say whether or not they are.
> No but then the point I'm making is we're drifting further and further away from Occam's razor.
Both sides of the debate have to explain the fact that a lot of developers disagree with them, so I don't think this argument really works.
> I do not let LLMs write code I care about, even with human review (because human review is imperfect).
That's fine but you are in the quickly vanishing minority.
> In what ways are current LLMs better excluding speed and cost (because getting those things by relaxing constraints on quality has always been trivially possible)? Even the fabled (heh) Mythos seems to be at best roughly equivalent to a competent human security researcher.
Well this is what I mean by benchmarks and measurement efforts. Lots of gaps in capabilities but we've had say superhuman competitive programming performance for awhile (including on fresh tasks not in training sets), extremely strong performance (super-p90-engineer) on say language-to-language porting, RE-bench (ML research engineering benchmark from METR) is already clearly above human perf, Mythos clearly (unless you believe this is all a massive fraud) has superior cyber capabilities, etc. Also, why do you discount speed and cost so much?
> Most of them aren't saying that the models are good enough to full replace developers, but this definitely isn't a strawman. I've been seeing the same basic claim for at least 18 months at this point.
Yea but what's the basic claim you're referring to here? Every model iteration is a significant bump up in performance according to a lot of complementary and principled measurements. What's been the thing that hasn't been true?
> Unless you have some non-LLM scores to compare to, those are still relative measures. They show/suggest that LLMs are getting better (at least in some ways), but without a definition of "good enough" in the same metric, that isn't sufficient to say whether or not they are.
There are human baselines in plenty of these benchmarks number one, and number two while no one is going to be able to tell you "once SWE-Bench Pro perf numbers get to X we can then refactor our existing process to completely offload task Y to agentic frameworks" thats a bit of a crazy ask. These numbers are pretty interpretable and many are pretty robust to things like training set leakage. What would you want to see here?
> Both sides of the debate have to explain the fact that a lot of developers disagree with them, so I don't think this argument really works.
Yet one side has a mountain of hard evidence and the other side has...an outdated n < 20 METR study using Sonnet 3?
> we've had say superhuman competitive programming performance for awhile
Fair. Question though, is this when compared to competitive programmers, or developers in general?
> extremely strong performance (super-p90-engineer) on say language-to-language porting
I'd need to see the methodology here and could easily be wrong, but I suspect this is largely down to "faster" and "willing to do a lot more of it without complaining"
> RE-bench (ML research engineering benchmark from METR) is already clearly above human perf
This pretty much has to be "relative to devs who don't specialize in that area", because if it wasn't the frontier labs wouldn't be paying a fortune to hire ML researchers.
> Mythos clearly has superior cyber capabilities
Based on Daniel Stenberg's experience with it [0], it seems like it's at best roughly on par with human experts. It's advantage is cost/speed.
> Also, why do you discount speed and cost so much?
Because in all the domains LLMs are applicable to, getting something cheaper/faster at the expense of quality isn't new or particularly interesting.
> Every model iteration is a significant bump up in performance according to a lot of complementary and principled measurements. What's been the thing that hasn't been true?
That they were good enough. To reuse the baby analogy, if every week your friend told you that their infant child was now heavier than an elephant (while acknowledging that the baby was lighter than one the previous week), and every week that turned out not to be true, it wouldn't be a defense of your friend to argue "ah, but the baby was heavier every week than the week before".
Also worth noting that as of ~8 months ago, while benchmark scores were steadily increasing, merge rates (aka whether the code was "good enough") were not [1].
> thats a bit of a crazy ask.
Why? If you use LLMs to do anything you're basically doing that already, it's just that the scope of your Y is smaller. Either the benchmarks are irrelevant and you're using something else to determine when that's appropriate for a given Y, or you do in fact have a value of X for the Y's you've handed over to LLMs.
> Yet one side has a mountain of hard evidence and the other side has...an outdated n < 20 METR study using Sonnet 3?
There's a lot of irony here, because by far the most common pro-LLM coding argument is "I feel like I'm producing good code faster with them", followed by "this other person feels like they're producing good code faster with them".
Also note that the most important part of the METR study you reference wasn't the slowdown they observed, it was the dramatic disagreement between what the participants thought the impact of AI was vs what it actually was. That isn't dependent on the model.
Why would they still need them if "[f]rontier models now outperform developers on many tasks, be it on quality/readability/maintainability, and let’s not talk about speed"
Because it’s a force multiplier at this stage in the capability ladder. ROI of developers is arguably higher or will become higher in a matter of months.
If AI could outperform humans, Anthropic would NEVER release that model. Instead, they'd use it to create a new google, photoshop, office, windows, etc for cheap then undercut all those companies and taking over the entire software industry.
It can outperform humans, just unevenly. You’ll see a lot of the same dynamics as you see with Mythos which, tbh is kind of refreshing. I get the sense that Dario while of course forced to ruthlessly run a company is genuinely interested in figuring out how to roll this out as ethically as he can.
> For coding you always want to go with the best model in the category [..]
And this is why many companies go out of business. You always want the best bang for your buck, sometimes this is the "best model" and sometimes it is not.
Cost for the value delivered. Like if you offered the current SOTA open source models at $0.1/M, I still think I'd be using Opus or 5.5 at $30/M. Or say GPT 5 which was released Aug 25, I don't think I'd use it for coding for even $0.1. I'd def find other uses for it(translations, agentic workflows, prompt guards etc), but for coding I don't think I'd ever completely switch to a SOTA open model
Unless ofc there was an actual speed difference, only reason I'd be willing to go with a worse model couple of percent worse than current best model is if the speed was at least 5x higher. Looking forward to kimi k2.6 offered publicly by Cerebras
Based on current market for LLMs I'd say my use of "you" in the general is fine. Even openrouter which doesn't capture all of the SOTA closed models but nearly all of opensource model usage has Opus as 1st(on last week) on "Programming" category and 3rd in overall rankings
I'd trust the OpenRouter rankings a lot more if they exposed the number of unique users for each model, as opposed to just a token count.
Currently I have no way of telling if big changes in their rankings are caused by a single "whale" switching providers, or if it's a more meaningful trend.
My point was that even openrouter, the one place people who are looking for open source SOTA models go to, doesn't definitively have opensource models at the top. Esp considering quite a lot of the closed models usage is through AWS, GCP , Azure etc, probably dwarfing the usage on openrouter by a huge factor
It's driven a lot by the harness too. If you're using claude code, you're actively being pushed towards newer models, even though older ones work perfectly fine for your use cases
Yeah wouldn't touch ClaudeCode when there are so many better harnesses that are free and portable. Seems like a waste of time to learn a proprietary tool when the FOSS ones are better.
>For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Currently, the difference is substantial, but what happens if capabilities saturate?
Then the house of cards comes crumbling down, but there is so much evidence to point to this not happening that it requires a bit of a theory for how that may happen
Well I think there are several fairly stable trends that paint a pretty compelling picture:
- performance scales with compute very very reliably. We have “scaling laws” (and have for years) and they are almost miraculously stable and show no sign of being invalidated at all even at the very largest scales. There are some theoretical bases for this though I’m not as familiar with the details
- these scaling laws are on an unintuitive quantity (validation loss on pretraining datasets), so we can look at downstream performance. Benchmarks are a minefield of junk but there are many decent ones and enough variety of techniques and data sources and scoring methods etc that in aggregate they are useful. The single number that I think is the best summary statistic across the crazy (O(100k)) number of benchmarks is the “epoch capability index” (just some branding over a reasonably standard statistical model that was really well thought out and a great idea). The trends in this are extremely stable. Eyeballing the trend over time on their graph we’re getting basically a GPT-4 to GPT-5 level capability improvement every ~18 months
- coding agents are not limited by the quality of the human training data they’re trained on, this is such a massive misconception: human data is only a bootstrap to a reinforcement learning phase. This combined with the fact that we have verifiable rewards means it’s just a matter of when not if for any given level of reliability.
- the massive compute investment implies that the compute that we’re building over the next 2-3 years will 10x the effective compute for training models. That combined with various R&D contributions (historically which have been very significant and there is no shortage of wins here), better data curation and flywheels, richer data (wait until conversation capability gets good) means we have several orders of magnitude of runway that we know of, today.
In short I don’t see any compelling evidence to suggest all of the trends we observe in many different ways will end any time soon.
I think RAG is a mostly outdated concept now, it's been subsumed by the idea of a "agent harness" which is exactly what Claude Code and Claude Cowork and OpenAI Codex and Claude.ai and ChatGPT themselves have now become.
An agent harness with access to a good search tool is a much more interesting thing than 2024-era RAG systems.
I appreciate where you are coming from, as you have surfed the front of the wave of GenAI for years. From my point of view, there is interesting because something is SOTA, and there is interesting because there is still more to build. I definitely understand state of RAG tech. I also view it as barely utilized versus what we can do with it, hence my question.
Agent harnesses integrated into good search tools are definitely interesting. Knowledgebasing with partitions and similar structure also remains fruitful for applications, above and beyond standard ElasticSearch on a cache.
> For coding you always want to go with the best model in the category
Will this always be true? There will never be an event horizon/point of diminishing returns where something not-bleeding-edge is "good enough" for 51%+ of users?
As long as closed source is 6 months ahead in terms of current difference. Although this is hard to figure out using simple percent based coding benchmarks, you def. notice it when you're actually trying to do a long task. Even simple things like UI "taste" is enough for me to use opus instead of 5.5 though even though 5.5 is strictly better for anything that doesn't have a UI, ie backend, scripts, making agent workflows etc
This is a silly take. There is a line of "good enough" for most coding (most CRUD apps and APIs are nothing special), and once we are past that, nobody will care about having the "newest, best" model except extreme outliers. And this base "good enough" model will become an ultra cheap commodity as we already see with GLM, deepseek, etc.
As long as closed models are 6 months ahead I won't be switching from them to prev. 6 month SOTA open source models. Maybe its just a different calculation if you're in a job, but as an indiehacker I'll take any edge I can get
Ofc again, can be convinced to switch if there's however a clear speed difference, like 5x+ for a open source sota even if it was SOTA for 6 months ago
I have stats from a harness that tells me glm5.1 is far more cost effective for us than Opus with the rate of defects and rework taken into account. In fact, with a decent harness I'm now increasingly favouring eHaiku over Opus for execution too. Opus is still worth it for planning, though, and far better at one-shotting things.
And you propose the same companies that have been cost cutting and avoiding buying you a chair for ever won't start objecting to a $200/dev/month subscription? The finance department won't have a say?
Oh, hey, I recognize you. Thank you for the very forward and thorough orbital sander recommendation at Home Depot. That's exactly what I wanted to deal with on my holiday weekend. You just know so much about this and the rest of us are simple passersbys.
Yep sorry was just pulling it out my rear, not like a market trend that nearly every enterprise uses Anthropic or Openai models for coding or that Anthropic has had such ridiculous growth that they're 10x-ing year over year
I'm ribbing you for writing like a condescending guru that invalidates the evaluatory capability of your peers. Not the meat of your evaluation (not to say that it's any good either, just that it's irrelevant).
If nothing else this blog did give me the idea that I should split my $200 claude max plan into two $100 CC max and $100 codex plan, esp because Claude is now offering 1.5x weekly limits so its the 5x usage is now more like 7.5x usage.
I really doubt that will happen anytime soon. I imagine there is still a great deal of preventative security-related work going on right now before they'll consider releasing this to the public (especially as this means granting access to Mythos to malicious actors).
I think its just a visual/default thing, cause Opus 4.0 isn't offered on claude code anymore. And opus 4.7 is on their official docs as a model you can change to, on claude code
Ik almost everyone is interested just in the SWE stuff, but this has been a good eval for me to think about how big the model is, how "creative" it is for generating new ideas etc.
More results from fable, with comparisons for Gemini, opus and some open source models: https://mesmer.tools/benchmarks/ai-video-generation
reply