Sign in Subscribe
news

2026-11-10 Opus 5.5 nerf claims - pin your harness, then measure

Sebastien Dubois

Sebastien Dubois

— 3 min read

Canonical version: 2026-11-10 Opus 5.5 nerf claims - pin your harness, then measure.

On 7 November, Shane Melaugh posted this on X: "is Opus 5.5 nerfed? I just had Gemini 4 Argon analyze 1,400 Claude Code sessions and it found that on November 3, something definitely happened. the model name didn't change, but the way it does reasoning took a sharp turn." The post got 582K views, and the Hacker News thread about it reached 783 points and about 480 comments.

Two days later, Anthropic released Claude Fable 5.2. For many people, that settled it. "Cue the new release", as one commenter put it.

In this piece, I'll go over what people reported, what the skeptics answered, and what I'd do before believing either side.

What the believers reported

Many commenters said they had noticed it too:

  • More mistakes, and repetitive errors in scripts that used to be rare (ghoul2)
  • Fumbled tool calls and unfinished tasks even at high, where low used to be enough (ricardobeat)
  • A Fable decline a couple of weeks earlier, then an Opus drop "even after bumping to Extra High". According to benwills, this comes back "about every couple months this year"
  • Token use that exploded: one user hit his 5-hour limit on a simple bug fix (shikck200)
  • A dip on Fable 5.1 the week before (PUSH_AX)

The theories: capacity moved to the next model, or a quantized model served before a launch (lukax); A/B tests (rzk, pointing to a Reddit thread); or the cynical version, from throwaw12, that LLMs are a "non-deterministic business" where the vendor can change anything and the customer can't prove it.

What the skeptics answered

  • "It's just a new summarizer, the model and performance itself is unchanged" (sunaookami)
  • orbital-decay explained why that matters: the reasoning text you see "is not real model's reasoning chain, they might have just replaced or tuned the model that summarizes it. If you want to track actual degradation you need to run evals"
  • LiveNerf, a nerf tracker on GitHub, showed no change. Its docs say: "Pin the CLI... a changed harness looks exactly like a changed model"
  • Claude Code updates change the system prompt, so the same model can behave differently from one week to the next (arecurrence)
  • The novelty of a model wears off, and people start noticing its flaws. That's a known bias (NewsaHackO)
  • The original analysis was itself done by a model, and nobody verified it (Topfi)

Not everyone was convinced by LiveNerf either. kingstnap argued it isn't sensitive enough to catch A/B tests, since those only hit some users.

As of today (10 November), there's no official answer from Anthropic that a web search turns up. And this isn't the first time: in September, another thread, "Livenerf: Has Opus 5.5 been nerfed yet?", got 394 comments.

My take

Vibes are not evals. A feeling that the model got worse can be right, but it can't tell apart a changed model, a changed summarizer, a changed system prompt, a changed effort setting, or a harder task than last week. And one model's analysis of 1,400 sessions is still one model's opinion.

I strongly believe the fix is simple: pin your harness, then measure.

  • Pin the Claude Code version you test with. Otherwise, you can't separate a model change from a harness change
  • Keep a small set of your own tasks with a known good result, and rerun them when something feels off. Same prompts, same effort level, same tools. See AI Evaluation
  • Log the model ID, the effort level and the token count of every run. With safeguard routing, the model that answered isn't always the one you selected
  • Compare the tokens per task too. "It got worse" and "it got more expensive" are two different problems, and the second one is easier to prove

The timing with Fable 5.2 proves nothing either way. Moving capacity before a launch is a plausible reason for a change, and the days before a launch are also when people look hardest for flaws.

One commenter (xscott) put it this way: "With open weights, YOU get to decide when the model changes." With a hosted model, you don't. The least you can do is keep the records that let you check. See AI Open Weight Models.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe