Opus 5.5 nerfed? Developers see a drop, benchmarks do not

On November 7, 2026, Shane Melaugh asked on X whether Opus 5.5 had been “nerfed”, meaning quietly made weaker. According to his analysis of 1,400 Claude Code sessions, the model changed the way it reasons around November 3. On Monday the same thread hit the front page of Hacker News (320 points and 215 comments within a few hours), and Anthropic has not commented so far.

What did developers actually notice?

Melaugh collected complaints from other developers: Opus 5.5 in Claude Code was said to ignore instructions, repeat the same mistakes and spin up subagents for trivial tasks. He then handed 1,400 of his own sessions to Gemini 4 Argon for analysis. The result: same model name, but around November 3 the way it reasons (quoting) “took a sharp turn”. The post is already close to 500,000 views.

It is not the only signal. On the Claude Code GitHub, issue #105112 has been open since October 26 with much harder numbers. Its author went through 6,638 Opus 5.5 responses from their transcripts and points to a different date: the switch between October 18 and 19.

1. Before the switch, they corrected 5.2% of 5,738 responses; after it, 25.4% of 900.
2. The increase shows up at every effort level: medium, high, xhigh and max.
3. The model is not thinking less: on average 4,434 thinking tokens per response in the week of October 12 and 6,702 a week later. According to the author, the problem is that the model does the research and then answers against what it just read.
4. In an update on November 5, they report 17.0% corrections on November 2–4, compared with 3.6–7.0% per week from late August to mid-October.

The author also checked that nothing changed on their side (same client version, CLAUDE.md, hooks). A second user with a similar pattern on Windows 11 joined the issue. Nobody from Anthropic has replied under it.

Opus 5.5 nerfed: what do the benchmarks say?

This is where the picture flips. Since launch, the LiveNerf project has run the same set of 78 questions (GPQA Diamond, MMLU-Pro, AIME and competition math) through Claude Code every day, graded without an AI judge. Status on November 8, day 46 of 60: “no change detected in Opus 5.5”. An earlier HN thread about LiveNerf gathered more than 900 points, so the “nerf” topic has kept coming back since launch.

But LiveNerf measures something different from what the complaining developers describe: single exam questions on a pinned Claude Code 2.1.280, not hours-long agentic sessions with tools. Both can be true at the same time.

Where could the difference come from?

First, thinking summaries. Claude does not show its raw reasoning, only a summary of it. Theo from T3 believes (we know this second-hand, from AICoder’s report) that Anthropic is gradually changing, account by account, exactly the module that writes those summaries, and that he sees no drop in capability. In that case, the Gemini analysis would be seeing a change in the style of the summaries, not in the model itself. Some HN commenters say the same.

Second, effort settings. According to Anthropic’s documentation, Opus 5.5 runs at medium by default, one level lower than Opus 5 (high). On top of that, recent Claude Code versions changed which effort subagents run with (Changelog, versions 2.1.316–2.1.317).

Third, precedent. As we know, in April Anthropic described in a postmortem that the complaints about Claude Code had three causes on the product side: a lowered default effort, a bug that cleared old thinking on every turn, and an instruction that shortened responses. Nobody touched the model weights back then, and the company stated that “we never intentionally degrade our models”. So a product change around the model can look like a “dumber model” even though it is not one.

What to watch for?

Until Anthropic speaks up, this is a dispute based on data from individual people, so let’s take it with a grain of salt. Still, issue #105112 is detailed enough (with an example request ID) that Anthropic can verify it on their side.

In our own setup, we can at least rule out the settings. According to the Changelog, in Claude Code you can set effort per project (effortLevel), for subagents through the CLAUDE_CODE_SUBAGENT_EFFORT variable, and “ultrathink” runs a turn on Opus 5.5 with xhigh effort. I wrote more about where Claude Code spends tokens in the posts on token usage from agent status updates and usage limits.

FAQ

Has Anthropic confirmed that Opus 5.5 got weaker?

No. As of November 9, I have not found any public comment from Anthropic, and nobody from the company has replied under issue #105112.

Has Anthropic ever weakened Claude before?

In April 2026, the company admitted that Claude Code quality dropped because of three product changes, fixed by version 2.1.116. According to the postmortem, the models themselves were not changed.

What is the default effort for Opus 5.5?

Medium. Opus 5 and earlier Opus models use high by default, so a request without an effort setting runs one level lower on Opus 5.5.