AI and security briefsModels

Is Claude Opus 5.5 nerfed since 3 November, and how do I check?

By Brandon Da Costa, founder of Digiton Dynamics

Published 9 November 2026Updated 9 November 2026Editorial policy

Nobody has confirmed a downgrade. A Hacker News thread of about 600 points says Opus 5.5 thinking changed on 3 November, and Anthropic has published no statement on it. Its status page lists no incident and its docs say a model ID never changes. Test your own workload instead.

Who should act: Teams that run Opus 5.5 in production or in Claude Code and want to know whether their own results moved, without relying on a thread of anecdotes.

What changed

The claim: a Hacker News submission titled "Is Opus 5.5 nerfed? Thinking patterns changed on November 3" reached about 600 points and 414 comments on 9 November (Hacker News, 9 November). It links to a post on X that our fetcher could not read, so we do not quote it. Commenters report worse work and fewer thinking blocks. One says a battle-style leaderboard put Opus 5.5 about 52 Elo lower after 3 November, with an interval from -160 to +40, and calls it inconclusive. Others see no change and blame the harness.

What Anthropic has published: its status page lists no incident about Opus 5.5 or output quality from 1 to 9 November. The only nearby entry is a Fable 5.1 error spike on 8 November (Anthropic status page, 9 November). Its API release notes for 2, 3 and 5 November do not mention Opus 5.5, thinking or effort (Anthropic release notes, 9 November). We found no Anthropic staff comment in the thread, though we read all but its last few thousand characters.

What the docs say: a model ID is a pinned version, and Anthropic does not update the weights or configuration of an existing ID. They also say the serving layer around it, including the router, safety classifiers and sampling logic, can change and cause minor differences in behaviour (Anthropic model ID docs, 9 November). Claude Code 2.1.317 (3 November) and 2.1.318 (4 November) changed subagent effort and dropped a resumed-thinking fix. The changelog does not link them to the reports (Claude Code changelog, 9 November).

Who it affects

• API users on claude-opus-5-5. By Anthropic's account the weights are fixed, so a real change would sit in the serving layer, your parameters or your prompts.
• Claude Code users. The harness is a separate moving part from the model, and its version changed on 3 and 4 November.
• Anyone who reads thinking text. Anthropic says a thinking block holds a summary, never the raw reasoning (Anthropic thinking docs, 9 November). A different summary is not proof of a different model.
• Anyone with a fallback. One commenter says Claude Code continued on Opus 4.8 after a refusal in 5 of 60 runs. The changelog lists a "Switch models when a message is flagged" setting.

What to do now

1. Freeze the inputs. Pin claude-opus-5-5, set effort explicitly (the default is medium), fix max_tokens and keep 30 to 50 real prompts in one file. Log claude --version for Claude Code.
2. Run every prompt five times and append one line per call.

   for i in 1 2 3 4 5
   do
     curl -s https://api.anthropic.com/v1/messages -H "x-api-key: $ANTHROPIC_API_KEY" -H "anthropic-version: 2023-06-01" -H "content-type: application/json" -d @request.json | jq -c '{day: (now|strftime("%F")), model, stop_reason, in: .usage.input_tokens, out: .usage.output_tokens}' >> runs.jsonl
   done

3. Compare days. Output tokens include thinking, so a shift in the median is the first thing to look for.

   jq -s 'group_by(.day)|map({day:.[0].day,n:length,median_out:(map(.out)|sort|.[length/2|floor])})' runs.jsonl

4. Score the answers too, with tests, exact-match fields or a fixed rubric. A change smaller than the spread between your five runs is noise.
5. Check the model field in each response and the stop_reason. A fallback or a truncated answer looks like a worse model.
6. Repeat weekly and after every client or harness update.

What to watch

• A statement from Anthropic staff or a status entry. Until one exists, treat a downgrade as unconfirmed.
• Public trackers. Commenters disagree about them and one says the methodology of the largest is unpublished, so use them as a hint and your own eval as the evidence.

Questions people ask

Digiton's production review checks that each service pins its model ID and settings and has a regression eval that would notice a change, and the checks are on the vibe coded app audit page.

How this brief was made: researched from the sources below, drafted with AI assistance and checked against each source on the dates shown. See the editorial policy.

Sources

1. Hacker News thread on the Opus 5.5 reports, 9 November 2026, read 9 November 2026. https://news.ycombinator.com/item?id=50422069
2. Anthropic status page, read 9 November 2026. https://status.claude.com
3. Anthropic API release notes, read 9 November 2026. https://platform.claude.com/docs/en/release-notes/overview
4. Anthropic docs on model IDs and versioning, read 9 November 2026. https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions
5. Anthropic docs on thinking, read 9 November 2026. https://platform.claude.com/docs/en/build-with-claude/thinking
6. Claude Code changelog, read 9 November 2026. https://code.claude.com/docs/en/changelog
7. Claude Opus 5.5 model page, 22 September 2026, read 9 November 2026. https://platform.claude.com/docs/en/models/opus-5-5/overview

More briefs

How these briefs are sourced and corrected. All AI and security briefs