I noticed Fable decline in quality a couple weeks ago. And I've noticed Opus drop over the last week, even after bumping to a default of Extra High reasoning for almost everything.
Tonight has been maddening with the kinds of absurd mistakes it's repeatedly making. It's so frustrating.
I'm the kind of person that wants something to be consistent and reliable. And Anthropic has been great, except for in this way. Which is enough of a recurring frustration (about every couple months this year) that I genuinely can't wait for other options to catch up. Then I'll be leaving as soon as I can.
I've been using Fable for a couple months as a relatively independent, long-horizon task executor and it worked well. About two weeks ago it completely stopped working. Like ~85% success rate down to ~10% in my workflows.
Totally possible that I'm just hallucinating/projecting/etc or my workflows somehow changes in terms of difficulty, but it's been kind of shocking.
I have the same exact experience. Working on one project, brainstorming. It was brilliant in July and August, like a seriously smarter, more experienced version of myself, reading in my mind, one step ahead of me, when pointed out a mistake it would remedy on the spot with the exact brilliant solution.
Now it's like talking to my hungover self, after 3 days bender, who just wanted to be left alone.
Definitely a significant change. I noticed it in just daily work. Quality of work down, mistakes up, repetitive errors in scripts - rare earlier, now a few every session.
You can still control it if you can trivially spin up a node with eight B200s and 2TB of memory to run your Deepseek v4 instance. Nobody on a subsidized plans is doing that. Local models that fit 128GB memory don't come close even to a "nerfed" frontier model.
> I noticed this too as did many others. Had to bump my default reasoning but it’s not the same. Seems to often happen before new model drops
I heard that Qwen4 was dropping some time soon, I wonder if Anthropic is trying to announce a new model immediately after Qwen4 - both for their IPO, and to kill the competition.
This has happened to consistently across OpenAI and Anthropic models that I can't believe people are still denying it. Obviously they're silently quantising their models or making other efficiency changes. If they can save 40% RAM and only hurt intelligence by a few percent in their benchmarks, they're going to do that. Unfortunately their benchmarks are obviously not representative of the real world, and these downgrades are very noticeable for people who rely on these models for work.
Of all the reasons to look at local inference, this is the one which pushes me in that direction the most. I want to rely on a model for consistency in my workflows. If I wake up on Monday and suddenly it's refactoring things it built on Friday, that's a huge problem. That makes it unreliable.
I guess the question really is: is this malice (trying to trick people into thinking new model is better by nerfing old one), or is this capacity constraints, where they obviously need to make room in their infra for inference of the new model by shoving the other models into a smaller box?
Eagerly awaiting Artificial Analysis to redo their Anthropic benchmarks and kick it down 15-20 index points, in the interests of accuracy. Those guys love accuracy. They moved Heaven and Earth to kick Kimi off the #2 spot, out of their deep, abiding passion for having the most accurate benchmarks in the industry.
I have no doubt they will ensure their users get an accurate picture of Opus 5.5 within hours.
I am using Opus 5.5 high almost exclusively. I definitely felt this and looked for nerfing watch sites. It felt like downgrading to the old version of deepseek. Things have improved a bit now.
I noticed Fable decline in quality a couple weeks ago. And I've noticed Opus drop over the last week, even after bumping to a default of Extra High reasoning for almost everything.
Tonight has been maddening with the kinds of absurd mistakes it's repeatedly making. It's so frustrating.
I'm the kind of person that wants something to be consistent and reliable. And Anthropic has been great, except for in this way. Which is enough of a recurring frustration (about every couple months this year) that I genuinely can't wait for other options to catch up. Then I'll be leaving as soon as I can.
I just don't know how long that will be.
I've been using Fable for a couple months as a relatively independent, long-horizon task executor and it worked well. About two weeks ago it completely stopped working. Like ~85% success rate down to ~10% in my workflows.
Totally possible that I'm just hallucinating/projecting/etc or my workflows somehow changes in terms of difficulty, but it's been kind of shocking.
I have the same exact experience. Working on one project, brainstorming. It was brilliant in July and August, like a seriously smarter, more experienced version of myself, reading in my mind, one step ahead of me, when pointed out a mistake it would remedy on the spot with the exact brilliant solution.
Now it's like talking to my hungover self, after 3 days bender, who just wanted to be left alone.
LiveNerf suggests not.
https://github.com/ninjahawk/livenerf
Token use exploded a week or two ago. I worked for a few hours and hit my 5 hour limit. Did only some simple bug fix. Somethings definitely broken.
This is the #1 thing i HATE with LLMs. They are not predictable like a compiler is.
It's not LLMs! Its their providers! If you self-hosted, you'd observe the slow down but you'd get same quality.
From what I understand they are picking and choosing who is running the latest models. So nerfed yes but for you not for everyone
Definitely a significant change. I noticed it in just daily work. Quality of work down, mistakes up, repetitive errors in scripts - rare earlier, now a few every session.
With open weights, YOU get to decide when the model changes.
Not true at all, there's much more to a model than its weights.
Like?
Thinking budgets…
In the context of this question, that's still something you as the user control.
Specifically
> YOU get to decide when the model changes
Yeah obviously but you can still control it.
You can still control it if you can trivially spin up a node with eight B200s and 2TB of memory to run your Deepseek v4 instance. Nobody on a subsidized plans is doing that. Local models that fit 128GB memory don't come close even to a "nerfed" frontier model.
As one user pointed out:
> I noticed this too as did many others. Had to bump my default reasoning but it’s not the same. Seems to often happen before new model drops
I heard that Qwen4 was dropping some time soon, I wonder if Anthropic is trying to announce a new model immediately after Qwen4 - both for their IPO, and to kill the competition.
This has happened to consistently across OpenAI and Anthropic models that I can't believe people are still denying it. Obviously they're silently quantising their models or making other efficiency changes. If they can save 40% RAM and only hurt intelligence by a few percent in their benchmarks, they're going to do that. Unfortunately their benchmarks are obviously not representative of the real world, and these downgrades are very noticeable for people who rely on these models for work.
Of all the reasons to look at local inference, this is the one which pushes me in that direction the most. I want to rely on a model for consistency in my workflows. If I wake up on Monday and suddenly it's refactoring things it built on Friday, that's a huge problem. That makes it unreliable.
I guess the question really is: is this malice (trying to trick people into thinking new model is better by nerfing old one), or is this capacity constraints, where they obviously need to make room in their infra for inference of the new model by shoving the other models into a smaller box?
Little of column A, little of column B.
These machines are inherently unreliable.
not as bad as sol 6.1. im getting better results out of glm 5.3
I was down voted into oblivion when I said that 3 weeks ago. Both Codex and Claude are absolutely nothing compared to what it was.
Eagerly awaiting Artificial Analysis to redo their Anthropic benchmarks and kick it down 15-20 index points, in the interests of accuracy. Those guys love accuracy. They moved Heaven and Earth to kick Kimi off the #2 spot, out of their deep, abiding passion for having the most accurate benchmarks in the industry.
I have no doubt they will ensure their users get an accurate picture of Opus 5.5 within hours.
I am using Opus 5.5 high almost exclusively. I definitely felt this and looked for nerfing watch sites. It felt like downgrading to the old version of deepseek. Things have improved a bit now.
It's dumbing down a little bit but I don't think its a lot worse. I think the change in display adds to the feel
Seems fine to me
It's just a new summarizer, the model and performance itself is umchamged.
+1, this tells nothing about the quality. Or even whether this is an upgrade or downgrade.
They're preparing to release a new model, so it has to look better than the current one.
Yes, I noticed it too, dumbing down continues, there goes their IPO.