I don't find this very compelling. If you look at the actual graph they are referencing but never showing [1] there is a clear improvement from Sonnet 3.7 -> Opus 4.0 -> Sonnet 4.5. This is just hidden in their graph because they are only looking at the number of PRs that are mergable with no human feedback whatsoever (a high standard even for humans).
And even if we were to agree that that's a reasonable standard, GPT 5 shouldn't be included. There is only one datapoint for all OpenAI models. That data point more indicative of the performance of OpenAI models (and the harness used) than of any progression. Once you exclude it it matches what you would expect from a logistic model. Improvements have slowed down, but not stopped
If you measure completion rate on a task where a single mistake can cause a failure, you won't see noticeable improvements on that metric until all potential sources of error are close to being eliminated, and then if they do get eliminated it causes a sudden large jump in performance.
That's fine if you just want to know whether the current state is good enough on your task of choice, but if you also want to predict future performance, you need to break it down into smaller components and track each of them individually.
I don't know that graph to me shows Sonnet 4.5 as worse than 3.7. Maybe the automated grader is finding code breakages in 3.7 and not breaking that out? But I'd much prefer to add code that is a different style to my codebase than code that breaks other code. But even ignoring that the pass rate is almost identical between the two models.
There is a decent case for this thesis to hold true especially if we look at the shift in training regimes and benchmarking over the last 1-2 years. Frontier labs don't seem to really push pure size/capability anymore, it's an all in focus on agentic AI which is mainly complex post-training regimes.
There are good reasons why they don't or can't do simple param upscaling anymore, but still, it makes me bearish on AGI since it's a slow, but massive shift in goal setting.
In practice this still doesn't mean 50 % of white collar can't be automated though.
> In practice this still doesn't mean 50 % of white collar can't be automated though.
Let me ask you this, though: if we wanted to, what percentage of white collar jobs could have been automated or eliminated prior to LLMs?
Meta has nearly 80k employees to basically run two websites and three mobile apps. There were 18k people working at LinkedIn! Many big tech companies are massive job programs with some product on the side. Administrative business partners, program managers, tech writers, "stewards", "champions", "advocates", 10-layer-deep reporting chains... engineers writing cafe menu apps and pet programming languages... a team working on in-house typefaces... the list goes on.
I can see AI producing shifts in the industry by reducing demand for meaningful work, but I doubt the outcome here is mass unemployment. There's an endless supply of bs jobs as long as the money is flowing.
I am pretty convinced that for most types of day to day work, any perceived improvements from the latest Claude models for example were total placebo. In blind tests and with normal tasks, people would probably have no idea if they're using Opus 4.5 or 4.6.
Interesting article, although with so few data points and such a specific time slice it is difficult to draw serious conclusions about the "improvement" of LLM models.
It's notably lacking newer models (4.5 Opus, 4.6 Sonnet) and models from Gemini.
LLMs appear to naturally progress in short leaps followed by longer plateaus, as breakthroughs are developed such as chain-of-thought, mixture-of-experts, sub-agents, etc.
Focusing on flashy breakthroughs hides the issue that bigger models and merge benchmarks rarely translate to reliability in real codebases. For routine merges, subtle regressions and context quirks matter more than headline progress. Unless evals stress nasty scenarios like multi-file renames with tricky conflicts, the numbers are mostly for show. Progress will plateau until someone tunes for the boring, messy cases that waste dev time.
That's an interesting claim, but I don't see it in my own work. They have got better but it's very hard to quantify. I just find myself editing their work much less these days (currently using GPT 5.4).
Without meaning to sound dismissive, because I'm really not intending to, there's also the possibility that you've gotten worse after enough time using them. You're treating yourself as a constant in this, but man cannot walk in the same river twice.
The problem with evals is the underlying rubric will always be either subjective, or a quantitative score based on something that is likely now baked into the training set directly.
You kind of have to go on "feels" for a lot of this.
I don't think it's true, but am I alone in wishing it was? My world is disrupted somewhat but so far I don't think we have a thing that upends our way of life completely yet. If it stayed exactly this good I'd be pretty content.
I agree with your sentiment, but I think we've yet to see the full application of the current technology. (Even if LLMs themselves don't improve, there's significant opportunity for people to use it in ways not currently being done)
I feel even if the models are stagnating, the tooling around them, and the integrations and harnesses they have are getting significantly more capable (if not always 'better' - the recent vscode update really handicapped them for some reason). Things like the new agent from booking.com or whatever, if it could integrate with all hotels, activities, mapping tools, flight system, etc could be hugely powerful.
Assuming we get no better than opus 4.6, they're very capable. Even if they make up nonsense 5% of the time!
1) Something happened during 2025 that made the models (or crucially, the wrapping terminal-based apps like Claude Code or Codex) much better. I only type in the terminal anymore.
2) The quality of the code is still quite often terrible. Quadruple-nested control flow abounds. Software architecture in rather small scopes is unsound. People say AI is āgood at front endā but I see the worst kind of atrocities there (a few days ago Codex 5.3 tried to inject a massive HTML element with a CSS before hack, rather than proprerly refactoring markup)
Two forces feel true simultaneously but in permanent tension. I still cannot make out my mind and see the synthesis in the dialectic, where this is truly going, if weāre meaningfully moving forward or mostly moving in circles.
I only say that because I'm a shit frontend dev. Honestly, I'm not that bad anymore, but I'm still shit, and the AI will probably generate better code than I will.
I think what happened with static image generation is happening with LLMs. Basically the tools around are becoming better, but all the AI improvements stall, the error rate stay the same (but external tools curate the results so it won't be noticeable if you don't run your own model), the accuracy is still slightly improving, but slower and slower, and never reach the 'perfect' point. Basically stablediffusion early 2025
Image quality has improved a lot in recent months thanks to better models. The ability of people to notice these improvements is plateauing because they are not trained to spot artifacts, which are becoming more obscure.
Controversial opinion from a casual user, but state-of-art LLMs now feel to me more intelligent then the average person on the steet. Also explains why training on more average-quality data (if there's any left) is not making improvements.
But LLMs are hamstrung by their harnesses. They are doing the equivalent of providing technical support via phone call: little to no context, and limited to a bidirectional stream of words (tokens). The best agent harnesses have the equivalent of vision-impairment accessibility interfaces, and even those are still subpar.
Heck, giving LLMs time to think was once a groundbreaking idea. Yesterday I saw Claude Code editing a file using shell redirects! It's barbaric.
I expect future improvements to come from harness improvements, especially around sub agents/context rollbacks (to work around the non-linear cost of context) and LLM-aligned "accessibility tools". That, or more synthetic training data.
entirely so. i think anthropic updated something about the compact algorithm recently, and its gone from working well over long times to basically garbage whenever a compact happens
Truth is I'm probably wrong. I should keep on testing ... but at the same time I precisely gave up because I didn't think the trend was fast enough to keep on investing on checking it so frequently. Now I just read this kind of post, ask around (mainly arguing with comments asking for genuine examples that should be "surprising" and kept on being disappointed) and that seems to be enough for a proxy.
I should though, as I mentioned in another comment, keep track of failed attempts.
PS: I check solely on self-hosted models (even if not on my machine but least on machines I could setup) because I do NOT trust the scaffolding around proprietary closed sources models. I can't verify that nobody is in the loop.
Well, on one hand they lack new data. Lot's of new code came out of an LLM, so it feeds back.
On the other hand, LLMs tend to go for an average by their nature (if you squint enough). What's more common in their training data, it's more common in the output, so getting them better without fundamental changes, requires one to improve the training data on average too which is hard.
What did improve a lot is the tooling around them. That's gotten way better.
Benchmaxxing aside, if you are using those tools for programming on a regular basis it should be self-evident that they are improving. I find it very hard to believe that someone using LLMs today vs what was available one year ago (Claude Code released Feb 2025) would have any difficulty answering this question.
I think it is important to try to find more rigorous things to test than the general sentiment of the people using the tools. If only because the more benchmarks we have the more we can improve models without regressions. METR is asking a really interesting question here, "are models improving at making one shot PRs?". The answer seems to be, yes, but slower than benchmarks suggest, if you look at the pass rate of different versions of Claude Sonnet. A reasonable answer is "you're not supposed to use them by making one shot PRs", but then ideally we would need to have some kind of standarized test for the ability of models to incorporate feedback and evolve PRs.
You really can't model these 5 data points with a linear regression or a step function. The models are of different sizes / use cases, and from two different labs. I feel like what we've observed generally is that different labs releasing similarly sized models at similar times are generally pretty similar.
I think the only reasonable thing to read into is Sonnet 3.5 -> 3.7 -> 4.5. But yeah, you just can't draw a line through this thing.
I will die on the hill that LLMs are getting better, particularly Anthropic's releases since December. But I can't point at a graph to prove that, I'm just drawing on my personal experience. I do use Claude Code though, so I think a large part of the improvement comes from the harness.
These studies are always really hard to judge the efficacy of. I would say though the most surprising thing to me about LLMs in the past year is how many people got hyped about the Opus 4.5 release. Having used Claude Code at work since it was released I haven't really noticed any step changes in improvement. Maybe that's because I've never tried to use it to one shot things?
Regardless I'm more inclined to believe that 4.5 was the point that people started using it after having given up on copy/pasting output in 2024. If you're going from chat to agentic level of interaction it's going to feel like a leap.
I used it with Sonnet 4.0 a lot, and there was vastly more back-and-forth and correction of "dumb" things, such as forgetting to add "using" statements in C# files.
I don't know if it's model, or harness improvements, or inbuilt-memory or all of the above, but it often has a step where it'll check itself that is done now before trying to build and getting an inevitable failure.
Those small things add up to a much smoother and richer experience today compared to 6 months ago.
My experience has been that raw āone-shot intelligenceā hasnāt improved as dramatically in the last year, but the workflow around the models has improved massively.
I've been able to supercharge a hobby project of mine over the last couple months using Opus 4.6 in claude code. I had to collaborate and write code still, but claude did like 75% of the work to add meaningful new features to an iOS/Android native mobile app, including Live Activities which is so overly complicated i would not have been able to figure that out. I have it running in a folder that contains both my back end api (express) and my mobile app (nativescript), so it does back end and front end work simultaneously to support new features. this wasnt possible 8 months ago.
I feel like anyone used AI coding tools before 11/25 and after 1/26 (with frontier models) will say there has been a massive jump in, there is a difference between whether LLM can do a specific task or pass some arguably arbitrary checks by maintainers vs. what the are capable of.
We still have tons of gaps about how to build and maintain code with AI, but LLM themselves getting better at an unbelievable pace, even with this kind of data analysis Iām surprised anyone can even question it.
It's my experience that opus 4, and then, particularly, 4.5, in Claude code, are head and shoulders above the competition.
I wrote an agentic coder years ago and it yielded trash. (Tried to make it do then what kiro does today).
The models are better. Now, caveat - I don't use anything but opus for coding - Sonnet doesn't do the trick. My experience with Codex and Gemini is that their top models are as good as Sonnet for coding...
I was trying to do something yestesrday and Claude was keep messing it up, after like an hour i realized the model somehow switched to sonet, opus 4.6 is crazy good. Itās very obvious in practice.
Although I feel like for chasing bugs and big systems codex is even better
This has been the general consensus for about three years now. "Drastic increases in capability have happened the last 3-6 months" have been a constant refrain.
Without any data from the study past September I think its not unreasonable, if you want to make an argument based on evidence.
For me personally, I agree with you, I'm really seeing it as well.
There's a consensus that SOMETHING changed with Opus 4.5. It might have been the "merge rates" metric, it might have not.
I'm certainly getting faster and cleaner-looking solutions for certain issues on Opus 4.6 than I was 5 months ago, but I'm not sure about the ability to solve (or even weigh in) the actual hard stuff, i.e. the stuff I'm paid for.
And I'm definitely not sure about the supposed big step between 4.5 and 4.6. I'm literally not seeing any.
I agree completely. I haven't noticed much improvement in coding ability in the last year. I'm using frontier models.
What's been the game changer are tools like Claude Code. Automatic agentic tool loops purpose built for coding. This is what I have seen as the impetus for mainstream adoption rather than noticeable improvements in ability.
I write a lot of C++ and QML code. Codex 5.3, only released in Feb, is the the first model I've used that would regularly generate code that passes my 25 years expert smell test and has turned generative coding from a timesap/nuisance into a tool I can somewhat rely on not to set me back.
Claude still wasn't quite there at the time, but I haven't tried 4.6 yet.
QML is a declarative-first markup language that is a superset of the JavaScript syntax. It's niche and doesn't have a giant amount of training data in the corpus. Codex 5.3 is the first model that doesn't super botch it or prefers to write reams of procedural JS embeds (yes, after steering). Much reduced is also the tendency to go overboard on spamming everything with clouds of helper functions/methods in both C++ and QML. It knows when to stop, so to speak, and is either trained or able to reason toward a more idiomatic ideal, with far less explicit instruction / AGENTS.md wrangling.
It's a huge difference. It might be the result of very specific optimization, or perhaps simultaneous advancements in the harness play a bigger role, but in my books my kneck of the woods (or place on the long tail) only really came online in 2026 as far as LLMs are concerned.
Maybe n=1, but I disagree? I notice that Sonnet 4.6 follows instructions much better than 4.5 and it generates code much closer to our already in-place production code.
It's just a point release and it isn't a significant upgrade in terms of features or capabilities, but it works... better for me.
From my personal experience, they have gotten better, but they havenāt unlocked any new capabilities. Theyāve just improved at what I was already using them for.
At the end of the day they still produce code that I need to manually review and fully understand before merging. Usually with a session of back-and-forth prompting or manual edits by me.
That was true 2 years ago, and itās true now (except 2 years ago I was copy/pasting from the browser chat window and we have some nicer IDE integration now).
Yeah I'm not buying the last bit about lower MSE with one term in the model vs two (Brier with one outcome category is MSE of the probabilities). That's the sort of thing that would make me go dig to find where I fucked up the calculation.
With one term it gets more robust in the face of excluding endpoints when constructing the jackknife train/test split, I think. But you're right, it does sound fishy.
Yesterday I asked a frontier model to help generate a report. It said great, it can do that, and output a table. I asked it to evaluate its prompt compliance in the result. It concluded that it had failed on every requirement. I asked why it had expressed such confidence, was it analagous to narcissism or psycopathy? It said no, and then said that if I just had to anthropomorphize it, I should think of it as a brilliant friend with severe frontal lobe brain damage.
>This means the step function has more predictive power (āfits betterā) than the linear slope. For fun, we can also fit a function that is completely constant across the entire timespan. That happens to get the best Brier score.
I mean, sure. but it's obvious in that graph that the single openai model is dragging down the right side. Wouldn't it be better to just stick to analyzing models from only one lab so that this was showing change over time rather than differences between models?
Even if one-shot LLM performance has plateaued (which I'm not convinced this data shows given omission of recent models that are widely claimed to be better) that missing the point that I see in my own work. The improved tooling and agent-based approaches that I'm using now make the LLM one-shot performance only a small part of the puzzle in terms of how AI tools have accelerated the time from idea to decent code. For instance the planning dialogs I now have with Claude are an important part of what's speeding things up for me. Also, the iterative use of AI to identify, track, and take care of small coding tasks (none of which are particularly challenging in terms of benchmarks) is simply more effective. Could this all have been done with the LLM engines of late 2024. Perhaps, but I think the fine-tuning (and conceivably the system prompts) that make the current LLM's more effective at agent-centered workflows (including tool-use) are a big part of it. One-shot task performance at challenging tasks is an interesting, certainly foundational, metric. But I don't think it captures the important advances I see in how LLM's have gotten better over the last year in ways that actually matter to me. I rarely have a well-defined programming challenge and the obligation to solve it in a single-shot.
> This means llms have not improved in their programming abilities for over a year. Isnāt that wild? Why is nobody talking about this?
Because it's not true. They have improved tremendously in the last year, but it looks like they've hit a wall in the last 3 months. Still seeing some improvements but mostly in skills and token use optimization.
> but mostly in skills and token use optimization.
I have heard rumors that token use optimization has been a recent focus to try to tidy up the financials of these companies before they IPO. take that with a grain of salt though
LLM's have 100% gotten better, but it's hard to say if it's "intrinsically better", if that makes sense.
> OpenAIās leading researchers have not completed a successful full-scale pre-training run that was broadly deployed for a new frontier model since GPT-4o in May 2024 [1]
That's evidence against "intrinsically better". They've also trained on the entire internet - we only have 1 internet, so.
However, late 2024 was the introduction of o1 and early 2025 was Deepseek R1 and o3. These were definitely significant reasoning models - the introduction of test time compute and significant RL pipelines were here.
Mid 2025 was when they really started getting integrated with tool calling.
Late 2025 is when they really started to become agentic and integrate with the CLI pretty well (at least for me). For example, codex would at least try and run some smoke tests for itself to test its code.
In early 2026, the trend now appears to be harness engineering - as opposed to "context engineering" in 2025, where we had to preciously babysit 1 model's context, we make it both easier to rebuild context (classic CS trick btw: rebooting is easier than restoring stale state [2]) and really lean into raw cli tool calling, subagents, etc.
FWIW, AI programming has still been as frustrating as it was when it was just TTC in 2025. Maybe because I don't have the "full harness" but it still has programming styles embedded such as silent fallback values, overly defensive programming, etc. which are obvoiusly gleaned from the desire to just pass all tests, rather than truly good programming design. I've been able to do more, but I have to review more slop... also the agents are really unpleasant to work with, if you're trying to have any reasonable conversation with them and not just delegate to them. It's as if they think the entire world revolves around them, and all information from the operator is BS, if you try and open a proper 2-way channel.
It seems like 2026 will go full zoom with AI tooling because the goal is to replace devs, but hopefully AI agents become actually nice to work with. Not sycophantic, but not passively aggressively arrogant either.
>To study how agent success on benchmark tasks relates to real-world usefulness, we had 4 active maintainers from 3 SWE-bench Verified repositories review 296 AI-generated pull requests (PRs). We had maintainers (hypothetically) accept or request changes for patches as well as provide the core reason they were requesting changes: core functionality failure, patch breaks other code or code quality issues.
I would also advise taking a look at the rejection reasons for the PRs. For example, Figure 5 shows two rejections for "code quality" because of (and I quote) "looks like a useless AI slop comment." This is something models still do, but that is also very easily fixable. I think in that case the issue is that the level of comment wanted hasn't been properly formalized in the repo and the model hasn't been able to deduce it from the context it had.
As for the article, I think mixing all models together doesn't make sense. For example, maybe a slope describe the increasing Claude Sonnet better than a step function.
Anecdotally, I haven't seen any real improvement from the AI tools I leverage. They're all good-ish at what they do, but all still lie occasionally, and all need babysitting.
I also wonder how much of the jump in early 2025 comes from cultural acceptance by devs, rather than an improvement in the tools themselves.
I think I'm coming to the same conclusion Gpt-3 to 5.3 have had real tangible but incremental improvements with quite diminishing returns.
Perhaps we won't see a phase change like improvement as we did from gpt-2 through to 3 until there is several more orders of magnitude parameters and/or training. Perhaps we will never see it again!
What is getting rapidly better is scaffolding but this seems to be more about understanding and building tools around LLMs than the LLMs themselves improving.
I'm still excited about AI but not constantly hyped to the rafters as some.
I think it depends on what you're using it for. If it is a simple kubernetes config then the model doesn't matter too much. Contract that with writing the scenario for a backtest for an algo that trades on a venue: it is not the same complexity and the basic models are terrible. I've had it tell me that it has added tests to find that they're just stubs!
Opus seems to be getting there, but on more complex tasks the others are a complete waste of time.
I don't find this very compelling. If you look at the actual graph they are referencing but never showing [1] there is a clear improvement from Sonnet 3.7 -> Opus 4.0 -> Sonnet 4.5. This is just hidden in their graph because they are only looking at the number of PRs that are mergable with no human feedback whatsoever (a high standard even for humans).
And even if we were to agree that that's a reasonable standard, GPT 5 shouldn't be included. There is only one datapoint for all OpenAI models. That data point more indicative of the performance of OpenAI models (and the harness used) than of any progression. Once you exclude it it matches what you would expect from a logistic model. Improvements have slowed down, but not stopped
1: https://metr.org/assets/images/many-swe-bench-passing-prs-wo...
Yes, I think this is basically an instance of the "emergent abilities mirage." https://arxiv.org/abs/2304.15004
If you measure completion rate on a task where a single mistake can cause a failure, you won't see noticeable improvements on that metric until all potential sources of error are close to being eliminated, and then if they do get eliminated it causes a sudden large jump in performance.
That's fine if you just want to know whether the current state is good enough on your task of choice, but if you also want to predict future performance, you need to break it down into smaller components and track each of them individually.
I don't know that graph to me shows Sonnet 4.5 as worse than 3.7. Maybe the automated grader is finding code breakages in 3.7 and not breaking that out? But I'd much prefer to add code that is a different style to my codebase than code that breaks other code. But even ignoring that the pass rate is almost identical between the two models.
There is a decent case for this thesis to hold true especially if we look at the shift in training regimes and benchmarking over the last 1-2 years. Frontier labs don't seem to really push pure size/capability anymore, it's an all in focus on agentic AI which is mainly complex post-training regimes.
There are good reasons why they don't or can't do simple param upscaling anymore, but still, it makes me bearish on AGI since it's a slow, but massive shift in goal setting.
In practice this still doesn't mean 50 % of white collar can't be automated though.
> In practice this still doesn't mean 50 % of white collar can't be automated though.
Let me ask you this, though: if we wanted to, what percentage of white collar jobs could have been automated or eliminated prior to LLMs?
Meta has nearly 80k employees to basically run two websites and three mobile apps. There were 18k people working at LinkedIn! Many big tech companies are massive job programs with some product on the side. Administrative business partners, program managers, tech writers, "stewards", "champions", "advocates", 10-layer-deep reporting chains... engineers writing cafe menu apps and pet programming languages... a team working on in-house typefaces... the list goes on.
I can see AI producing shifts in the industry by reducing demand for meaningful work, but I doubt the outcome here is mass unemployment. There's an endless supply of bs jobs as long as the money is flowing.
I am pretty convinced that for most types of day to day work, any perceived improvements from the latest Claude models for example were total placebo. In blind tests and with normal tasks, people would probably have no idea if they're using Opus 4.5 or 4.6.
It's because they are getting so good it's impossible to recognize them.
Haiku 4.5 is already so good it's ok for 80% (95%?) of dev tasks.
I'd agree with you on 4.5 to 4.6, but going from gpt-5 or 4.0 to 4.5 was night and day.
Interesting article, although with so few data points and such a specific time slice it is difficult to draw serious conclusions about the "improvement" of LLM models.
It's notably lacking newer models (4.5 Opus, 4.6 Sonnet) and models from Gemini.
LLMs appear to naturally progress in short leaps followed by longer plateaus, as breakthroughs are developed such as chain-of-thought, mixture-of-experts, sub-agents, etc.
Focusing on flashy breakthroughs hides the issue that bigger models and merge benchmarks rarely translate to reliability in real codebases. For routine merges, subtle regressions and context quirks matter more than headline progress. Unless evals stress nasty scenarios like multi-file renames with tricky conflicts, the numbers are mostly for show. Progress will plateau until someone tunes for the boring, messy cases that waste dev time.
That's an interesting claim, but I don't see it in my own work. They have got better but it's very hard to quantify. I just find myself editing their work much less these days (currently using GPT 5.4).
Without meaning to sound dismissive, because I'm really not intending to, there's also the possibility that you've gotten worse after enough time using them. You're treating yourself as a constant in this, but man cannot walk in the same river twice.
The problem with evals is the underlying rubric will always be either subjective, or a quantitative score based on something that is likely now baked into the training set directly.
You kind of have to go on "feels" for a lot of this.
I don't think it's true, but am I alone in wishing it was? My world is disrupted somewhat but so far I don't think we have a thing that upends our way of life completely yet. If it stayed exactly this good I'd be pretty content.
I agree with your sentiment, but I think we've yet to see the full application of the current technology. (Even if LLMs themselves don't improve, there's significant opportunity for people to use it in ways not currently being done)
I feel even if the models are stagnating, the tooling around them, and the integrations and harnesses they have are getting significantly more capable (if not always 'better' - the recent vscode update really handicapped them for some reason). Things like the new agent from booking.com or whatever, if it could integrate with all hotels, activities, mapping tools, flight system, etc could be hugely powerful.
Assuming we get no better than opus 4.6, they're very capable. Even if they make up nonsense 5% of the time!
I feel that two things are true at the same time:
1) Something happened during 2025 that made the models (or crucially, the wrapping terminal-based apps like Claude Code or Codex) much better. I only type in the terminal anymore.
2) The quality of the code is still quite often terrible. Quadruple-nested control flow abounds. Software architecture in rather small scopes is unsound. People say AI is āgood at front endā but I see the worst kind of atrocities there (a few days ago Codex 5.3 tried to inject a massive HTML element with a CSS before hack, rather than proprerly refactoring markup)
Two forces feel true simultaneously but in permanent tension. I still cannot make out my mind and see the synthesis in the dialectic, where this is truly going, if weāre meaningfully moving forward or mostly moving in circles.
> People say AI is āgood at front endā
I only say that because I'm a shit frontend dev. Honestly, I'm not that bad anymore, but I'm still shit, and the AI will probably generate better code than I will.
The models lose the ability to inject subtle and nuance stuff as they scale up, is what Iāve observed.
>fischer warned us against eyeballing plots proceeds to eyeball it with an arbitrary function
There was a long flat line before the step, models improve, but PR pass rate without human intervention is inherently a staircase function
I think what happened with static image generation is happening with LLMs. Basically the tools around are becoming better, but all the AI improvements stall, the error rate stay the same (but external tools curate the results so it won't be noticeable if you don't run your own model), the accuracy is still slightly improving, but slower and slower, and never reach the 'perfect' point. Basically stablediffusion early 2025
Image quality has improved a lot in recent months thanks to better models. The ability of people to notice these improvements is plateauing because they are not trained to spot artifacts, which are becoming more obscure.
Controversial opinion from a casual user, but state-of-art LLMs now feel to me more intelligent then the average person on the steet. Also explains why training on more average-quality data (if there's any left) is not making improvements.
But LLMs are hamstrung by their harnesses. They are doing the equivalent of providing technical support via phone call: little to no context, and limited to a bidirectional stream of words (tokens). The best agent harnesses have the equivalent of vision-impairment accessibility interfaces, and even those are still subpar.
Heck, giving LLMs time to think was once a groundbreaking idea. Yesterday I saw Claude Code editing a file using shell redirects! It's barbaric.
I expect future improvements to come from harness improvements, especially around sub agents/context rollbacks (to work around the non-linear cost of context) and LLM-aligned "accessibility tools". That, or more synthetic training data.
> But LLMs are hamstrung by their harnesses
entirely so. i think anthropic updated something about the compact algorithm recently, and its gone from working well over long times to basically garbage whenever a compact happens
Steet? Do you mean street? They're smarter in the same way a search engine is smarter.
I gave up on trying months ago, you can see the timeline on top of https://fabien.benetou.fr/Content/SelfHostingArtificialIntel...
Truth is I'm probably wrong. I should keep on testing ... but at the same time I precisely gave up because I didn't think the trend was fast enough to keep on investing on checking it so frequently. Now I just read this kind of post, ask around (mainly arguing with comments asking for genuine examples that should be "surprising" and kept on being disappointed) and that seems to be enough for a proxy.
I should though, as I mentioned in another comment, keep track of failed attempts.
PS: I check solely on self-hosted models (even if not on my machine but least on machines I could setup) because I do NOT trust the scaffolding around proprietary closed sources models. I can't verify that nobody is in the loop.
Well, on one hand they lack new data. Lot's of new code came out of an LLM, so it feeds back.
On the other hand, LLMs tend to go for an average by their nature (if you squint enough). What's more common in their training data, it's more common in the output, so getting them better without fundamental changes, requires one to improve the training data on average too which is hard.
What did improve a lot is the tooling around them. That's gotten way better.
Benchmaxxing aside, if you are using those tools for programming on a regular basis it should be self-evident that they are improving. I find it very hard to believe that someone using LLMs today vs what was available one year ago (Claude Code released Feb 2025) would have any difficulty answering this question.
I think it is important to try to find more rigorous things to test than the general sentiment of the people using the tools. If only because the more benchmarks we have the more we can improve models without regressions. METR is asking a really interesting question here, "are models improving at making one shot PRs?". The answer seems to be, yes, but slower than benchmarks suggest, if you look at the pass rate of different versions of Claude Sonnet. A reasonable answer is "you're not supposed to use them by making one shot PRs", but then ideally we would need to have some kind of standarized test for the ability of models to incorporate feedback and evolve PRs.
You really can't model these 5 data points with a linear regression or a step function. The models are of different sizes / use cases, and from two different labs. I feel like what we've observed generally is that different labs releasing similarly sized models at similar times are generally pretty similar.
I think the only reasonable thing to read into is Sonnet 3.5 -> 3.7 -> 4.5. But yeah, you just can't draw a line through this thing.
I will die on the hill that LLMs are getting better, particularly Anthropic's releases since December. But I can't point at a graph to prove that, I'm just drawing on my personal experience. I do use Claude Code though, so I think a large part of the improvement comes from the harness.
These studies are always really hard to judge the efficacy of. I would say though the most surprising thing to me about LLMs in the past year is how many people got hyped about the Opus 4.5 release. Having used Claude Code at work since it was released I haven't really noticed any step changes in improvement. Maybe that's because I've never tried to use it to one shot things?
Regardless I'm more inclined to believe that 4.5 was the point that people started using it after having given up on copy/pasting output in 2024. If you're going from chat to agentic level of interaction it's going to feel like a leap.
I used it with Sonnet 4.0 a lot, and there was vastly more back-and-forth and correction of "dumb" things, such as forgetting to add "using" statements in C# files.
I don't know if it's model, or harness improvements, or inbuilt-memory or all of the above, but it often has a step where it'll check itself that is done now before trying to build and getting an inevitable failure.
Those small things add up to a much smoother and richer experience today compared to 6 months ago.
Nah, pre 4.5 it was not comfortable to use agentic coding.
My experience has been that raw āone-shot intelligenceā hasnāt improved as dramatically in the last year, but the workflow around the models has improved massively.
When you combine models with:
tool use
planning loops
agents that break tasks into smaller pieces
persistent context / repos
the practical capability jump is huge.
I've been able to supercharge a hobby project of mine over the last couple months using Opus 4.6 in claude code. I had to collaborate and write code still, but claude did like 75% of the work to add meaningful new features to an iOS/Android native mobile app, including Live Activities which is so overly complicated i would not have been able to figure that out. I have it running in a folder that contains both my back end api (express) and my mobile app (nativescript), so it does back end and front end work simultaneously to support new features. this wasnt possible 8 months ago.
I feel like anyone used AI coding tools before 11/25 and after 1/26 (with frontier models) will say there has been a massive jump in, there is a difference between whether LLM can do a specific task or pass some arguably arbitrary checks by maintainers vs. what the are capable of.
We still have tons of gaps about how to build and maintain code with AI, but LLM themselves getting better at an unbelievable pace, even with this kind of data analysis Iām surprised anyone can even question it.
Data is missing on this chart.
It's my experience that opus 4, and then, particularly, 4.5, in Claude code, are head and shoulders above the competition.
I wrote an agentic coder years ago and it yielded trash. (Tried to make it do then what kiro does today).
The models are better. Now, caveat - I don't use anything but opus for coding - Sonnet doesn't do the trick. My experience with Codex and Gemini is that their top models are as good as Sonnet for coding...
I was trying to do something yestesrday and Claude was keep messing it up, after like an hour i realized the model somehow switched to sonet, opus 4.6 is crazy good. Itās very obvious in practice.
Although I feel like for chasing bugs and big systems codex is even better
Given that it is the general consensus that a step function occurred with Opus 4.5/4.6 only 3 months ago - it seems like an insane omission.
This has been the general consensus for about three years now. "Drastic increases in capability have happened the last 3-6 months" have been a constant refrain.
Without any data from the study past September I think its not unreasonable, if you want to make an argument based on evidence.
For me personally, I agree with you, I'm really seeing it as well.
There's a consensus that SOMETHING changed with Opus 4.5. It might have been the "merge rates" metric, it might have not.
I'm certainly getting faster and cleaner-looking solutions for certain issues on Opus 4.6 than I was 5 months ago, but I'm not sure about the ability to solve (or even weigh in) the actual hard stuff, i.e. the stuff I'm paid for.
And I'm definitely not sure about the supposed big step between 4.5 and 4.6. I'm literally not seeing any.
I had this suspicion for a while I think we just got way better in harnessing not the models actual reasoning
So we got better in giving it the right context and tools to do the stuff we need to do but not the actual thinking improvements
I agree completely. I haven't noticed much improvement in coding ability in the last year. I'm using frontier models.
What's been the game changer are tools like Claude Code. Automatic agentic tool loops purpose built for coding. This is what I have seen as the impetus for mainstream adoption rather than noticeable improvements in ability.
My anecdotal experience is rather different.
I write a lot of C++ and QML code. Codex 5.3, only released in Feb, is the the first model I've used that would regularly generate code that passes my 25 years expert smell test and has turned generative coding from a timesap/nuisance into a tool I can somewhat rely on not to set me back.
Claude still wasn't quite there at the time, but I haven't tried 4.6 yet.
QML is a declarative-first markup language that is a superset of the JavaScript syntax. It's niche and doesn't have a giant amount of training data in the corpus. Codex 5.3 is the first model that doesn't super botch it or prefers to write reams of procedural JS embeds (yes, after steering). Much reduced is also the tendency to go overboard on spamming everything with clouds of helper functions/methods in both C++ and QML. It knows when to stop, so to speak, and is either trained or able to reason toward a more idiomatic ideal, with far less explicit instruction / AGENTS.md wrangling.
It's a huge difference. It might be the result of very specific optimization, or perhaps simultaneous advancements in the harness play a bigger role, but in my books my kneck of the woods (or place on the long tail) only really came online in 2026 as far as LLMs are concerned.
Maybe n=1, but I disagree? I notice that Sonnet 4.6 follows instructions much better than 4.5 and it generates code much closer to our already in-place production code.
It's just a point release and it isn't a significant upgrade in terms of features or capabilities, but it works... better for me.
From my personal experience, they have gotten better, but they havenāt unlocked any new capabilities. Theyāve just improved at what I was already using them for.
At the end of the day they still produce code that I need to manually review and fully understand before merging. Usually with a session of back-and-forth prompting or manual edits by me.
That was true 2 years ago, and itās true now (except 2 years ago I was copy/pasting from the browser chat window and we have some nicer IDE integration now).
As they become more capable peoples commits will also become more ambitious.
So Iād say fairly flat commit acceptance numbers make sense even in the context of improving LLMs
Indeed. Why is this post down voted? Thereās always trade-offs taking place, itās good to call them out.
This means llms have not improved in their programming abilities for over a year. Isnāt that wild? Why is nobody talking about this?
Because hype makes money.
Yeah I'm not buying the last bit about lower MSE with one term in the model vs two (Brier with one outcome category is MSE of the probabilities). That's the sort of thing that would make me go dig to find where I fucked up the calculation.
With one term it gets more robust in the face of excluding endpoints when constructing the jackknife train/test split, I think. But you're right, it does sound fishy.
Yesterday I asked a frontier model to help generate a report. It said great, it can do that, and output a table. I asked it to evaluate its prompt compliance in the result. It concluded that it had failed on every requirement. I asked why it had expressed such confidence, was it analagous to narcissism or psycopathy? It said no, and then said that if I just had to anthropomorphize it, I should think of it as a brilliant friend with severe frontal lobe brain damage.
That actually helps.
If you look at a separate trend for the smaller Sonnet models, you can see a rapid trend
3.7 to 4.5 looks pretty flat here.
>This means the step function has more predictive power (āfits betterā) than the linear slope. For fun, we can also fit a function that is completely constant across the entire timespan. That happens to get the best Brier score.
I mean, sure. but it's obvious in that graph that the single openai model is dragging down the right side. Wouldn't it be better to just stick to analyzing models from only one lab so that this was showing change over time rather than differences between models?
Even if one-shot LLM performance has plateaued (which I'm not convinced this data shows given omission of recent models that are widely claimed to be better) that missing the point that I see in my own work. The improved tooling and agent-based approaches that I'm using now make the LLM one-shot performance only a small part of the puzzle in terms of how AI tools have accelerated the time from idea to decent code. For instance the planning dialogs I now have with Claude are an important part of what's speeding things up for me. Also, the iterative use of AI to identify, track, and take care of small coding tasks (none of which are particularly challenging in terms of benchmarks) is simply more effective. Could this all have been done with the LLM engines of late 2024. Perhaps, but I think the fine-tuning (and conceivably the system prompts) that make the current LLM's more effective at agent-centered workflows (including tool-use) are a big part of it. One-shot task performance at challenging tasks is an interesting, certainly foundational, metric. But I don't think it captures the important advances I see in how LLM's have gotten better over the last year in ways that actually matter to me. I rarely have a well-defined programming challenge and the obligation to solve it in a single-shot.
No Gemini. No Opus 4.5. No GPT codex.
As they said, ragebait used to be believable.
They are getting better, but they are also hitting diminishing returns.
There's only so much data to train on, and we are unlikely to see giant leaps in performance as we did in 2023/2024.
2026-27 will be the years of primarily ecosystem/agentic improvements and reducing costs.
> This means llms have not improved in their programming abilities for over a year. Isnāt that wild? Why is nobody talking about this?
Because it's not true. They have improved tremendously in the last year, but it looks like they've hit a wall in the last 3 months. Still seeing some improvements but mostly in skills and token use optimization.
> but mostly in skills and token use optimization.
I have heard rumors that token use optimization has been a recent focus to try to tidy up the financials of these companies before they IPO. take that with a grain of salt though
After only 3 months (!) you can claim a plateau, but not a wall.
How the "costant function" result fits the data points better than a slope that has two parameters instead of one.
Cross-validation. The slope overfits when the test set is included from the data the model is fitted on.
In my niche the Opus 4.6 has been a game changer. In comparison all other LLMs look stupid. I am considering cancelling all other subscriptions.
LLM's have 100% gotten better, but it's hard to say if it's "intrinsically better", if that makes sense.
> OpenAIās leading researchers have not completed a successful full-scale pre-training run that was broadly deployed for a new frontier model since GPT-4o in May 2024 [1]
That's evidence against "intrinsically better". They've also trained on the entire internet - we only have 1 internet, so.
However, late 2024 was the introduction of o1 and early 2025 was Deepseek R1 and o3. These were definitely significant reasoning models - the introduction of test time compute and significant RL pipelines were here.
Mid 2025 was when they really started getting integrated with tool calling.
Late 2025 is when they really started to become agentic and integrate with the CLI pretty well (at least for me). For example, codex would at least try and run some smoke tests for itself to test its code.
In early 2026, the trend now appears to be harness engineering - as opposed to "context engineering" in 2025, where we had to preciously babysit 1 model's context, we make it both easier to rebuild context (classic CS trick btw: rebooting is easier than restoring stale state [2]) and really lean into raw cli tool calling, subagents, etc.
[1] https://newsletter.semianalysis.com/p/tpuv7-google-takes-a-s...
[2] https://en.wikipedia.org/wiki/Kernel_panic
FWIW, AI programming has still been as frustrating as it was when it was just TTC in 2025. Maybe because I don't have the "full harness" but it still has programming styles embedded such as silent fallback values, overly defensive programming, etc. which are obvoiusly gleaned from the desire to just pass all tests, rather than truly good programming design. I've been able to do more, but I have to review more slop... also the agents are really unpleasant to work with, if you're trying to have any reasonable conversation with them and not just delegate to them. It's as if they think the entire world revolves around them, and all information from the operator is BS, if you try and open a proper 2-way channel.
It seems like 2026 will go full zoom with AI tooling because the goal is to replace devs, but hopefully AI agents become actually nice to work with. Not sycophantic, but not passively aggressively arrogant either.
From the METR study (https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs...):
>To study how agent success on benchmark tasks relates to real-world usefulness, we had 4 active maintainers from 3 SWE-bench Verified repositories review 296 AI-generated pull requests (PRs). We had maintainers (hypothetically) accept or request changes for patches as well as provide the core reason they were requesting changes: core functionality failure, patch breaks other code or code quality issues.
I would also advise taking a look at the rejection reasons for the PRs. For example, Figure 5 shows two rejections for "code quality" because of (and I quote) "looks like a useless AI slop comment." This is something models still do, but that is also very easily fixable. I think in that case the issue is that the level of comment wanted hasn't been properly formalized in the repo and the model hasn't been able to deduce it from the context it had.
As for the article, I think mixing all models together doesn't make sense. For example, maybe a slope describe the increasing Claude Sonnet better than a step function.
Anecdotally, I haven't seen any real improvement from the AI tools I leverage. They're all good-ish at what they do, but all still lie occasionally, and all need babysitting.
I also wonder how much of the jump in early 2025 comes from cultural acceptance by devs, rather than an improvement in the tools themselves.
I think I'm coming to the same conclusion Gpt-3 to 5.3 have had real tangible but incremental improvements with quite diminishing returns.
Perhaps we won't see a phase change like improvement as we did from gpt-2 through to 3 until there is several more orders of magnitude parameters and/or training. Perhaps we will never see it again!
What is getting rapidly better is scaffolding but this seems to be more about understanding and building tools around LLMs than the LLMs themselves improving.
I'm still excited about AI but not constantly hyped to the rafters as some.
I think it depends on what you're using it for. If it is a simple kubernetes config then the model doesn't matter too much. Contract that with writing the scenario for a backtest for an algo that trades on a venue: it is not the same complexity and the basic models are terrible. I've had it tell me that it has added tests to find that they're just stubs! Opus seems to be getting there, but on more complex tasks the others are a complete waste of time.
Itās better pre and post training + better harnessing