GPT-5 usage is 20% higher on days that start with "S"

Nevertheless, 7 datapoints does not a trend make (and the data presented certainly doesnt explain why). The daily variation is more than I would have expected, but could also be down to what day of the week the pizza party is or the weekly scrum meetings is at a few of their customers workplaces.

For development use cases, I switched to Sonnet 4.5 and haven't looked back. I mean, sure, sometimes I also use GPT-5 (and mini) and Gemini 2.5 Pro (and Flash), and also Cerebras Code just switched to providing GLM 4.6 instead of the previous Qwen3 Coder so those as well, but in general the frontier models are pretty good for development and I wouldn't have much reason to use something like Sonnet 4 or 3.7 or whatever.

I wish we could pin down not only the model but also the way the UI works as well.

Last week Claude seemed to have a shift in the way it works. The way it summarises and outputs its results is different. For me it's gotten worse. Slower, worse results, more confusing narrowing down what actually changed etc etc.

Long story short, I wish I was able to checkpoint the entire system and just revert to how it was previously. I feel like it had gotten to a stage where I felt pretty satisfied, and whatever got changed ... I just want it reverted!

We tried GPT-5 for a RAG use case, and found that it performs worse than 4.1. We reverted and didn't look back.

I've found that the VSCode GitHub Copilot extension defaults to Claude Sonnet 4.0 (in agent mode) in all new workspaces. It's the first thing I check now, but I imagine a lot of people just roll with it, especially if they use inline completions where it might not be obvious what model is being used.

Just one week data right after the release, when it is already one month later?

This data is basically meaningless, show us the latest stats.

Seems to completely ignore usage of local/free models as well as anything but Sonnet/ChatGPT. So my confidence in the good faith of the author is... heavily restricted.

> Each model appears to emphasize a different balance between reasoning and execution. Rather than seeking one “best” system, developers are assembling model alloys—ensembles that select the cognitive style best suited to a task.

This (as well as the table above it) matches my experience. Sonnet 4.0 answers SO-type questions very fast and mostly accurately (if not on a niche topic), Sonnet 4.5 is a little bit more clever but can err on the side of complexity for complexity's sake, and can have a hard time getting out of a hole it dug for itself.

ChatGPT 5 is excellent at finding sources on the web; Gemini simply makes stuff up and continues to do so even when told to verify; ChatGPT provides link that work and are generally relevant.

Even for non-developer use cases o3 is a much better model for me than GPT5 on any setting.

30 seconds-1 minute is just the time I am patient enough to wait as that's the time I am spending on writing a question.

Faster models just make too many mistakes / don't understand the question.

I use both Codex and Claude, mostly cuz it's cheaper to jump between them than to buy a Max sub for my use-case. My subjective experience is that Codex is better with larger or weird, speghetti-ish codebases, or codebases with more abstract concepts, while Claude is good for more direct uses. I haven't spent significant time fine-tuning the tools for my codebases.

Once, I set up a proxy that allowed Claude and Codex to "pair program" and collaborate, and it was cool to watch them talk to each other, delegate tasks, and handle different bits and pieces until the task was done.

It could be an interesting data point, but without correcting for absolute usage figures and their customers it's kind of hard to make general statements.

I don't get the point of this post. Personally, I think that the thinking process is essential for accurate tool usage. Whenever I interact with Claude family models, either on a web chat or via a coding agent CLI, I believe that this thinking process is what makes Claude more accurate in using tools.

It could be true that newer models just produce more tokens seemingly out of no reasons. But with the increasing number of tool definitions, in the long run, I think it will pay off.

Just a few days ago, I read "Interleaved Thinking Unlocks Reliable MiniMax-M2 Agentic Capability"[1]. I think they have a valid point that this thinking process has significance as we are moving towards agents.

[1] https://www.minimax.io/news/why-is-interleaved-thinking-impo...

Tangential to this: what are the most reliable benchmarks for LLM in coding these days?

My team still uses Sonnet 3.5 for pretty much everything we do because it's largely enough and it's much, much faster than newer models. The only reason we're switching is because the models are getting deprecated...

The usage data looks like a classic case of the drift principle. When a model gets heavily optimized for alignment, polish, and safety, you gain consistency but lose some fidelity to the actual task. Newer models think longer, act less, and smooth over edges that used to be useful for real work. Older models aren’t smarter, they’re just sitting earlier on the drift curve, before over compression starts eroding decisiveness. So the specialization we’re seeing may just be developers picking the version where fidelity holds up best, not the one with the highest benchmark score.

Some missing context (pun intended) is that Augment code has recently switched to a per-token instead of per-message pricing model. This hasn't gone down particularly well, but that's another story. But it may well be that users drop back to older models in the expectation it will use less tokens.

Personally, I stopped using GPT-5 as it would just be tool call after tool call without ever stopping to tell you what the hell it was doing. Sonnet 4.5 much better in this regard. Albeit it's too verbose for the new token based world ('let me just summarise that in a report')

Multiple models is a must, mostly due to the sometimes unpredictable variations in responses to specific situations/contexts/languages and frameworks. I find that Sonnet 4, Gemini Pro 2.5 are solid in comparison to newer models (especially Sonnet 4.5 which I find frequently to underperform). When one model is stuck in a loop, switching to a model like GPT-5 often breaks it but which model will work is subject to circumstance. P.S. I spend at least 3-4 hours a day in code-gen activities of various levels using Cursor as my primary IDE.

Curious that this omits Opus.

Opus 4.1 beats Sonnet 4.5 and Codex for me still in any coding tasks. In planning it's slighly behind Codex but only slightly.

Caveat: I do almost exclusively Rust (computer graphics).

I think this is somewhat disingenuous since not everyone uses the latest thing, and people tend to stick to “what works” for them.

Models are picky enough about prompting styles that changing to a new model every week/month becomes an added chunk of cognitive overload, testing and experimentation, plus even in developer tooling there have been minor grating changes in API invocations and use of parameters like temperature (I have a fairly low-level wrapper for OpenAI, and I had to tweak the JSON handling for GPT-5).

Also, there are just too many variations in API endpoints, providers, etc. We don’t really have a uniform standard. Since I don’t use “just” OpenAI, every single tool I try out requires me to jump through a bunch of hoops to grab a new API key, specify an endpoint, etc.—and it just gets worse if you use a non-mainstream AI endpoint.

I think this is one of the many indicators that even though these models get “version upgrades” it’s closer to switching to a different brain that may or may not understand or process things the way you like. Without a clear jump in performance, people test new models and move back to ones they know work if the new ones aren’t better or are actually worse.

Not complaining too loudly because improvement is magical, but trying to stay on top of model cards and knowing which one to use for specific cases is bit tedious.

I think the end game is decent local model that does 80% of the work, and that also knows when to call the cloud, and which models to call.

Matches my experience too. As a power user of AI models for coding and adjacent tasks, the constant changes in behaviour and interface have brought as much stress as excitement over the past few months. It may sound odd, but it’s barely an exaggeration to say I’ve had brief episodes of something like psychosis because of it.

For me, the “watering down” began with Sonnet 4 and GPT-4o. I think we were at peak capability when we had:

- Sonnet 3.7 (with thinking) – best all-purpose model for code and reasoning

- Sonnet 3.5 – unmatched at pattern matching

- GPT-4 – most versatile overall

- GPT-4.5 – most human-like, intuitive writing model

- O3 – pure reasoning

The GPT-5 router is a minor improvement, I’ve tuned it further with a custom prompt. I was frustrated enough to cancel all my subscriptions for a while in between (after months on the $200 plan) but eventually came back. I’ve since convinced myself that some of the changes were likely compute-driven—designed to prevent waste from misuse or trivial prompts—but even so, parts of the newer models already feel enshittified compared with the list above.

A few differences I've found in particular:

- Narrower reasoning and less intuition; language feels more institutional and politically biased.

- Weaker grasp of non-idiomatic English.

- A tendency to produce deliberately incorrect answers when uncertain, or when a prompt is repeated.

- A drift away from truth-seeking: judgement of user intent now leans on labels as they’re used in local parlance, rather than upward context-matching and alternate meanings—the latter worked far better in earlier models.

- A new fondness for flowery adjectives. Sonnet 3.7 never told me my code was “production-ready” or “beautiful.” Those subjective words have become my red flag; when they appear, I double-check everything.

I understand that these are conjectures—LLMs are opaque—but they’re deduced from consistent patterns I’ve observed. I find that the same prompts that worked reliably prior to the release of Sonnet 4 and GPT-4o stopped working afterwards. Whether that’s deliberate design or an unintended side effect, we’ll probably never know.

I usually switch models depending on the situation, for simpler stuff, I lean toward 4o since it’s faster to get answers.

But when things get more complex, I prefer GPT-5, talking with it often gives me fresh ideas and new perspectives.

Isn’t this obvious? When you have a task you think is hard. You give it to a cleverer model. When a task is straight forward you give it to an older one.

Doesn't surprise me. davinci-002 was better than davinci-003. The core breakthrough has been done, stuff's just shifting around now.

GPT5 is HELLISHLY slow. That's all there is to it.

It loves doing a whole bunch of reasoning steps and prolaim how mucf of a very good job it did clearing up its own todo steps and all that mumbo jumbo, but at the end of the day, I only asked it a small piece of information about nginx try_files that even GPT3 could answer instantly.

Maybe before you make reasoning models that go on funny little sidequests wher they multiply numbers by 0 a couple of times, make it so its good at identfying the length of a task. ntil then, I'll ask little bro and advance only if necessity arrives. And if it ends up gathering dust, well... yeah.

I'm surprised they don't mention cost or latency, would imagine that would be a factor as well.

To those who complain about GPT5 being slow; I recently migrated https://app.sqlai.ai and found that setting service_tier = “priority” makes it reason twice as fast.

I am definitely not. Claude 4.5 and GPT 5 all the way for me

This is how the bubble pops.

I've been thinking the AI bubble wouldn't pop, because even the AI advances we've already seen can change the majority of industries if it is carefully integrated with existing technology. But if there's a mass movement to use older and/or smaller models, then yeah, all the money going into newer bigger models will pop.

Or, maybe the training datasets getting polluted with AI slop will mean that new models are worse than old models. That would pop the industry.

Or, maybe the GPT-4 era was the golden era for AI, and making them bigger and better is just overfitting (in the classical machine learning sense of the word) and is both worse and more expensive. This would pop the industry too.

I guess there's a few ways for the industry to pop, but this trend of using older models makes me especially skeptical of AI.

grok-code-fast-1 is my current pick, found accuracy and speed better than Sonnet 4.5 for day-to-day usage.

> At Augment Code, we run multiple frontier models side by side in production.

I mean, this is technically false, right? They’re not running these models but calling the APIs? Not that it matters.

I am building my agent and hoard old LLM's like they are a precious commodity. Older models are less censored, more flavorful and don't have that RL slop factor. Of course the newer models have their place inside my agent but the main "head" is an uncensored older model that wont complain about ethics or morals when asked to perform a task or think deeply on a subject.

To the authors of the site, please know that your current "Cookiebot by Usercentrics" is old and pretty much illegal. You shouldn't need to click 5 times to "Reject all" if accepting all is one click. Newer versions have a "Deny" button.

[flagged]