The idea is that new models can handle up to a certain level of complexity, at which point they fall apart. Every new model can handle more complexity, so there's a wonderful time upon release when you feel like you can do anything, only for you to hit the complexity ceiling a few days later when you saturate it. Rinse and repeat for the next model.
I have not been doing increasingly complex things since Opus 4.6 when models got really good.
My work at my job has stayed the same. But the model quality has varied.
They definitely tune the models in production after launch, if not only to share load during high traffic times. It’s not a crazy conspiracy that the same model can be stupider at different times.
> I have not been doing increasingly complex things since Opus 4.6 when models got really good.
This is a more a statement on the work you do and how you work versus the models. I'm doing more complex work since Fable (and now for way cheaper thanks to Opus 5.5)
With 4.6 I would still babysit a lot more code quality and so on. With the newer model I see myself talking about features at a higher level, and then not having to nitpick PRs to death. Which means most of my time is now spent talking to the model about the product instead of the implementation of the product.
It's not about doing more complex things - complexity is more dictated by how large your codebase is, etc.
> It’s not a crazy conspiracy that the same model can be stupider
Sorry, I really do think it's a conspiracy. If nerfing were real, it would be trivial to prove. DeepSWE, SWEBench, and other benchmarks are all available for anyone to run. A "nerfing" hypothesis has to survive the fact that a statistically significant dip in benchmarks has never been observed.
Nerf is real, i think we initially get full precision models and later quants. My own logs show it clearly for opus 4.5 to 5, consistently a few months post launch, models start making quant based mistakes, like slipping in inappropriate tokens (e.g. chinese ones in english text) which doesnt happen at all in the first few months and regularly later. Additionally frontier problems previously done well start being done poorly, until later model variants where performance mostly holds, likely due to them training on your data reguardless of what boxes you tick.
My local models don't display that degradation, sensed or measured. They consistently perform equally to what I expect of them, precisely because they don't change.
How does twitter explain that? Is my internal model for expectation of capacity magically not drifting for local models but somehow is for anthropic api call based models?
They test it on launch day, then benchmark it against that. A deviation of above 10% is considered a change. They're currently tracking Opus 5.5 and GPT-6 Astra.
This bench famously detected a degradation of Opus 4.6 which Anthropic later blogged about. I personally think people sense nerfs more often than they happen and that it's often about honeymoon effects.
I dunno, I never sense nerfs for local models, but consistently a few months after launch for corpo hosted models, seems odd my internal model for the capacity of a model drifts for anthropic models but not local ones.
The only reason why claude fable is better than opus in my opinion is that it has more "criteria"... if you present a problem and then ask for his recommendation you can get an opinion on why and reasoning on why that one... Opus is going to vomit 10k lines of extremely dense prose in nerdify++ level.
Yesterday I fought claude fable to not just jump to make changes like a dog following a treat, that we were researching... at some point I introduced the word HAWAI... and only if I say HAWAI the thing can start making changes..
I was going to post here in HN just to have a "I knew this was the reason" when they release fable > 5.1
I had the exact same feeling every time they have a new big release
Anecdata: I've been running a long-lived claude code session with Opus 4.6 for the last few days. Yesterday, almost right after the Sonnet 5.5 announcement, codex starting asking for permission to run things a lot more often
The quality of the output/work seems the same, but the speed at which it gets stuff done is a lot slower, because it's asking for permission so much more
I don't have any numbers/stats, just my impression. However, I imagine that if Anthropic could make the models ask for permission more often, it could be an interesting way to throttle access, without degrading quality of the output
I wonder if more organizations approving the model on a fast-tracked basis means Anthropic is straining for more compute and thus sheds a tiny bit to handle the increased demand, especially at peak times.
(which has led me to believe that's a good approximation for hedonic adaptation, I've seen tons of attempts at demonstrating nerfing via benches, none persist)
I wonder if API is affected by this issue, especially Claude on public clouds? Would that means the subsidized rate just means they use cheaper quantized models and it's not comparable to API spending.
I've always used Enterprise per-token billing for Claude Code and I've never understood these nerf complaints. I've never noticed any slow downs at certain times of day, or a gradual decline in quality.
There’s probably contractual guarantees in the enterprise plans. My understanding of the subscriptions is they can swap the models out if any of them is getting too heavily loaded for a period of time
It seems as if this is based on demand. Whenever a new model is released, I'm guessing tens of thousands of us switch over to try the latest and greatest, which overloads the servers, leading to nerfing. It's 100% dishonest, but they realized they would lose users a lot quicker if they were honest and just said "our models are overloaded, come back later".
After Fable launch I switched over to Codex and it was simply amazing, with frequent usage resets that seemed never ending. They clearly had more compute than they knew what to do with. Post Astra, Codex has gotten dumb again across all models, increased usage for no real reason, and no resets.
I'm guessing Opus 5.5 will take the heat off Codex for a bit, leading to better performance. So I guess I stick around here instead of switching again?
All this dishonesty and shadiness is part of why open models feel inevitable. Even if the total cost of ownership is higher (debatable; seems that way at small scales, but likely not as you grow), I'd rather have intelligence controlled by me that works for me.
The current period is as pro-customer as we're ever going to get, with cash still flying around and neither OpenAI nor Anthropic on the public market, and people are already forced into this sort of business to keep them true to their word. The point isn't even whether they're nerfing the models (I don't think they are), but that people can't seem to trust them to do right.
Theory (Conjecture? Hypothesis?): What we notice as "model nerfing" is the company diverting compute to training/running new unreleased models..
Remember that some people get access to the next flagship version long before us peasants do. I recall seeing the mention of "Astra" more than a month before it was officially announced
Has there ever been any measurement of this, of any sort? Honest question. I frequently see a plural of anecdotes to that effect, but I've not seen a concrete statement of fact or measurement that could be scrutinized or tested in any way.
If so, please share. This should be measurable, and I'm glad this project is measuring it.
Answers in the form of additional anecdotes, stated with even greater passion but still lacking a statement that could be tested and falsified, would validate my exact concern.
The claim was that there's an experience of a model losing power. Your claim amounts to "No, you are not experiencing what you say". That's quite a claim for you to make with no data and no argument.
I was just wondering if, like certain processors, bugs get fixed and the speed goes down. Like, they find it's doing things it shouldn't, restrict it, and harm the throughput.
Yeh it's absurd that people claim this all the time. It's some crazy conspiracy theory and when you ask for examples nothing ever shows up.
It would be economical suicide from anthropic and OpenAI to actually need models intentionally.
But hey I guess it's hard with technology that truly seems like magic.
People say if you'd bring electricity to the middle ages you'd be called a witch and burned. The same is happening to the model labs here because they are bringing tech that the world isn't ready for yet.
I made a graphic to explain why people feel like the models get nerfed:
https://x.com/thesilenceturns/status/2103551351825543610
The idea is that new models can handle up to a certain level of complexity, at which point they fall apart. Every new model can handle more complexity, so there's a wonderful time upon release when you feel like you can do anything, only for you to hit the complexity ceiling a few days later when you saturate it. Rinse and repeat for the next model.
Btw, you have a typo in the twitter handle on your profile (not in your comment), 'thesilencesturns'.
Anthropic has admitted to nerfing in the past. On top of that, model performance changes as they move compute to schwaggier providers as well.
Your chart is wrong.
Where?
My work at my job has stayed the same. But the model quality has varied.
They definitely tune the models in production after launch, if not only to share load during high traffic times. It’s not a crazy conspiracy that the same model can be stupider at different times.
This is a more a statement on the work you do and how you work versus the models. I'm doing more complex work since Fable (and now for way cheaper thanks to Opus 5.5)
With 4.6 I would still babysit a lot more code quality and so on. With the newer model I see myself talking about features at a higher level, and then not having to nitpick PRs to death. Which means most of my time is now spent talking to the model about the product instead of the implementation of the product.
> It’s not a crazy conspiracy that the same model can be stupider
Sorry, I really do think it's a conspiracy. If nerfing were real, it would be trivial to prove. DeepSWE, SWEBench, and other benchmarks are all available for anyone to run. A "nerfing" hypothesis has to survive the fact that a statistically significant dip in benchmarks has never been observed.
My local models don't display that degradation, sensed or measured. They consistently perform equally to what I expect of them, precisely because they don't change.
How does twitter explain that? Is my internal model for expectation of capacity magically not drifting for local models but somehow is for anthropic api call based models?
https://www.bridgebench.ai/nerf-bench
They test it on launch day, then benchmark it against that. A deviation of above 10% is considered a change. They're currently tracking Opus 5.5 and GPT-6 Astra.
This bench famously detected a degradation of Opus 4.6 which Anthropic later blogged about. I personally think people sense nerfs more often than they happen and that it's often about honeymoon effects.
Yesterday I fought claude fable to not just jump to make changes like a dog following a treat, that we were researching... at some point I introduced the word HAWAI... and only if I say HAWAI the thing can start making changes..
I was going to post here in HN just to have a "I knew this was the reason" when they release fable > 5.1
I had the exact same feeling every time they have a new big release
The quality of the output/work seems the same, but the speed at which it gets stuff done is a lot slower, because it's asking for permission so much more
I don't have any numbers/stats, just my impression. However, I imagine that if Anthropic could make the models ask for permission more often, it could be an interesting way to throttle access, without degrading quality of the output
(which has led me to believe that's a good approximation for hedonic adaptation, I've seen tons of attempts at demonstrating nerfing via benches, none persist)
After Fable launch I switched over to Codex and it was simply amazing, with frequent usage resets that seemed never ending. They clearly had more compute than they knew what to do with. Post Astra, Codex has gotten dumb again across all models, increased usage for no real reason, and no resets.
I'm guessing Opus 5.5 will take the heat off Codex for a bit, leading to better performance. So I guess I stick around here instead of switching again?
The current period is as pro-customer as we're ever going to get, with cash still flying around and neither OpenAI nor Anthropic on the public market, and people are already forced into this sort of business to keep them true to their word. The point isn't even whether they're nerfing the models (I don't think they are), but that people can't seem to trust them to do right.
Open models are the endgame.
The classifier is a model; it examines the actual prompt.
They already do this for the safety "guardrails".
Remember that some people get access to the next flagship version long before us peasants do. I recall seeing the mention of "Astra" more than a month before it was officially announced
If so, please share. This should be measurable, and I'm glad this project is measuring it.
Answers in the form of additional anecdotes, stated with even greater passion but still lacking a statement that could be tested and falsified, would validate my exact concern.
It would be economical suicide from anthropic and OpenAI to actually need models intentionally.
But hey I guess it's hard with technology that truly seems like magic. People say if you'd bring electricity to the middle ages you'd be called a witch and burned. The same is happening to the model labs here because they are bringing tech that the world isn't ready for yet.