I noticed the same at work using Opus 5.5 on medium which was the sweet spot. It’s now just not great.

For me Claude was nerfed when:

- They put the shock collar on it [Cyber] Your session has been paused - meaning I have to start a new session. That is also on Opus 4.8 now, just not quite as sensitive as on 5.5. (Defensive coding, defensive system configuration) In my opinion Claude is not permitted to write secure code or analyze ones existing code for bugs that may be security related.

- They downgraded my 20x account to a 10x, so I downgraded it to the lowest one.

- I keep hitting context memory limits. Claude is not proactively compacting when it should, especially on max. It says max does not matter but in my experience it very much does.

I have not logged in for a couple weeks so I have no idea if the model itself has been nerfed. The feeling I got and it's just intuition is that they want to focus on enterprise customers and would prefer individuals went away so I cleared out my data and moved my projects elsewhere. Message received.

Claude's security restrictions seem more like a CYA move than anything related to security. In particular, restrictions on analyzing artifacts uploaded or supplied by public or legitimate authenticated access to erxternal sites seem counterproductive as far as security is concerned.

I find the verification requirements particularly puzzling: low level requirements that amount to "use competing LLMs to spam reports to build cred", high level requirements that preclude detailed analysis of anything that flys below the radar of large actors.

The 20x was always 10x. The 20x only applied to 5-hour limits which are now gone (which is good). This was communicated

Why not share a distributed benchmark? Tokens are shared across volunteers; everyone gets informed of the nerf simultaneously. No one need be surprised.

It seems like pre-IPO companies need to balance "cost" vs. "optics" and it's one of the metrics they optimize.

So, launch at an "exceptional" state and then cut costs after some time following the release, then repeat to make it appear like lots of advancement is happening.

I have zero evidence. But anecdotally (and working with Claude every day) I see wild swings in Opus/Fables abilities day to day. It really does feel like they are twiddling knobs behind the scenes and hoping no one notices.

They are actually doing it, it's the best business method you can do. Who says it's not, shut the fuck up you idiot, anthropic implemented a secret safeguard without announcement, of course they can tweak it

Oh man, can't wait for open-source to catch up...

> We trust them that what is behind the string "opus-5-5" is consistent and what we signed up for.

Do you run evals as a matter of course? Everyone I know that does shows high levels of variance across them under the best case.

That’s not to say that the labs don’t downgrade models at times, just that without any measurement it’s a non-interesting assertion.

I've noticed this too, since about Wednesday.

I doubt Anthropic are quantizing their own models. Since effort levels are essentially a compute budget I wonder if they’re using this behind the scenes to dial it down when getting close to capacity. Thinking blocks are hidden so we wouldn’t be able to tell right?

I guess a new model release is right around the corner

Polymarket says 80% chance of fable 5.2 release over the next 8 days, it feels pretty consistent to me with other release periods, where the existing models feel "dumbed down".

funnily livenerf is showing the model to be slightly more intelligent these days: https://github.com/ninjahawk/livenerf

Livenerf is also using a subscription to run, not API, so it's apparently not a subscription thing. Maybe if you're unlucky you're served with a quantized model until they release fable 5.2 ?

You have to assume they detect benchmarking / testing and feed those clients consistent models.

Fable 5.2 has been released:

Claude Fable 5.2 was officially launched on November 9, 2026.

I have not noticed this myself, but it could easily happen... I guess until we have regulation to prevent this that we can actually enforce (in other words: never) there's not much we can do about it. Except let the market do its job, and go to other AI companies... which we won't, since we're locked in and Anthropic is using its market leader position.

Hardly a new business practice I guess, but deceptive nevertheless.