Is Claude Opus 5.5 nerfed? As of November 9, 2026, there is no confirmed nerf. The viral November 3 chart tracks the visible thinking text, which Anthropic says is written by a separate summarizer model, so it cannot prove the model itself got worse, and an independent daily tracker shows no accuracy drop on a fixed question set. One detailed report of answers getting worse since October 19 is still open with no reply from Anthropic, so the honest answer is "probably not on November 3, unresolved for October 19."

This guide walks through what each piece of evidence actually measures, what Anthropic has said before about quality complaints, and what you can do if Claude feels worse for your own work.

Why people are asking whether Claude Opus 5.5 is nerfed

Claude Opus 5.5 launched on September 22, 2026, priced at $4 per million input tokens and $20 per million output tokens, according to Anthropic's launch post and its model overview. Anthropic pitched it as a cheaper flagship that performs at the level of Claude Fable 5.1 on most work. For how it compares with the rest of the lineup, see Claude Fable vs Opus vs Mythos.

Complaints that Opus 5.5 got "dumber" have come in waves since launch. The latest started on November 7, when developer Shane Melaugh posted on X that he had Gemini 4 Argon analyze 1,400 of his Claude Code sessions and that "something definitely happened" on November 3, even though the model name did not change. The claim was covered by LavX News and sysnote, and a Hacker News thread on it passed 250 points on November 9, with commenters split between "it is obviously worse" and "show me the data."

Google autocomplete reflects the same split: "is claude nerfed", "is claude getting worse" and "is opus 5.5 nerfed" all return long lists of suggestions this week.

The four things "nerfed" can mean

"The model got worse" can point to four different layers, and the evidence for each is very different.

LayerWhat would changeWho can see it
Model weightsA different or compressed model behind the same nameOnly Anthropic
Serving and routingHardware, sampling, safety fallbacks, server promptsMostly Anthropic
The app or harnessDefault effort, system prompts, how old thinking is keptPartly users, through release notes
Your own workHarder tasks, longer sessions, different habitsYou

Labs say the first layer does not change silently. Anthropic wrote in April that "we never intentionally degrade our models." Most past quality problems sat in the second and third layers, which is why a fair answer has to look at all four.

What the November 3 chart actually measures

Melaugh's chart tracks the wording of the thinking text shown in Claude Code: which phrases thinking blocks start with, and how often certain patterns appear. Those patterns shifted sharply around 1 PM ET on November 3, in the middle of active sessions.

The catch is in Anthropic's own thinking documentation. It says what you see "is never the raw chain of thought" but a summary, and that "summarization is processed by a different model from the one you target in your requests." On Opus 5.5 the API hides thinking text by default (the display setting defaults to "omitted"), and you are billed for the full thinking tokens either way.

So a change in how thinking reads is, first of all, a change in the summarizer's writing. T3 founder Theo Browne checked his own logs and reported that the new summary style reached his accounts on different days between late October and November 4, with reasoning volume and output speed roughly unchanged. He concluded that Anthropic appears to be rolling out a new reasoning summarizer account by account, and wrote that he saw "no sign of degraded capability," as reported by sysnote and Tech Twitter.

Melaugh also posted a second chart claiming less thinking before the first tool call at high effort. His data and method have not been published, so it cannot be checked yet. We found no Anthropic announcement of a summarizer change.

The October 19 report that is still open

The stronger evidence for a real change is older. In GitHub issue #105112 on Anthropic's Claude Code repository, one heavy user scored 6,638 Opus 5.5 answers from local transcripts. A script flagged answers the user had to correct, or that Claude retracted in its next reply.

  • Before October 19: 5.2% of 5,738 answers were corrected.
  • From October 19 to 25: 25.4% of 900 answers were corrected.
  • The user dates the step to between 8:52 PM PT on October 18 and 1:13 PM PT on October 19, with the same client build, instructions and cached prompt prefix on both sides.
  • A November 4 update put the weekly rate at 23.0%, 20.5% and then 17.0%, against 3.6% to 7.0% from late August to mid October.

The failure described is not laziness. Tool calls per answer held steady and billed thinking rose, but the model "reads my files and runs the lookups... and then writes an answer that ignores or contradicts what it read." Raising effort did not fix it. A second user filed a similar report (#105477) on Windows 11 with a different Claude Code build.

This is still one person's workflow, measured by a script, and the author admits closer checking after early misses could inflate later days. But it is specific, dated and reproduced by a second user. As of November 9, the issue's comments come only from users, which i10x and sysnote also note.

What the independent tracker shows

LiveNerf is an open source project that runs the same 78 questions through Opus 5.5 every day on a Claude Max subscription, with a pinned Claude Code version, built on the UK AI Security Institute's Inspect framework. Its November 8 update (day 46) reports no change across three full ten day windows after launch week, with the latest window slightly above baseline.

Two limits matter. The project says it can only detect accuracy shifts of about 6.6 points per window, and in validation it could not reliably tell Opus 5.5 apart from Opus 5. It also asks single questions with known answers, with no tools or files. The October 19 complaint is about long sessions where the model ignores files it just read, which this tracker does not test. So the tracker rules out a big drop in raw problem solving, not every kind of regression.

Known reasons Opus 5.5 can feel weaker

Some "it feels worse" reports have documented causes that are not a secret downgrade.

Lower default effort. Anthropic's effort documentation says Opus 5.5 defaults to medium, while Opus 5 and earlier Opus models default to high, "so a request that omits effort runs one level lower." Anthropic's Opus 5.5 prompting guide says that in its testing Opus 5.5 at medium matches or exceeds Opus 5 at high on coding and knowledge work, but your tasks may differ.

Thinking is always on and mostly hidden. Adaptive thinking cannot be turned off on Opus 5.5, so effort is the main lever. Because the text you see is a summary, a worse summary can make good reasoning look sloppy.

Safety fallbacks. The launch post says Opus 5.5's cybersecurity, biology and distillation safeguards "fall back to another model transparently," with most cybersecurity tasks rerouted to Opus 4.8. That is documented behavior for flagged requests, and there is no public evidence it touches ordinary work. Hidden safety behavior is a live research topic; see our explainer on GPT-6 Astra covert underperformance for a different case.

Anthropic has been here before

On April 23, 2026, Anthropic published an update on Claude Code quality reports after weeks of complaints. It traced them to three product changes, not the model or the API:

  1. On March 4 it lowered Claude Code's default effort from high to medium, then reverted on April 7.
  2. On March 26 it shipped a change meant to clear old thinking once after an hour idle; a bug made it clear every turn, fixed April 10.
  3. On April 16 it added a system prompt line to cut verbosity, which hurt coding quality and was reverted April 20.

Anthropic said the API and inference layer were unaffected, that the reports were hard to tell apart from normal variation at first, and it reset usage limits for subscribers. The lesson for today: real regressions have happened, they were in the product layer, and they were confirmed only after weeks of user reports.

Pros and cons of each explanation

ExplanationFor itAgainst it
Weights were swapped or compressedMany users feel a dropNo tracker sees it; Anthropic denies doing this
Summarizer changed on Nov 3Mid session switch; account by account rollout in one user's logsNot confirmed by Anthropic
Something changed around Oct 19Dated, detailed report plus a second oneOne workflow; no reply yet
Effort default and harnessDocumented medium defaultExplains a constant gap, not a sudden one
Noise and harder tasksRun to run variance is realDoes not fit a sharp, dated step

What to do if Claude feels worse for you

  • Set effort explicitly. If you use the Anthropic API, pass high or xhigh rather than relying on the medium default, and compare results on the same task.
  • Keep a small private test set. Five or ten tasks from your real work, rerun a few times, will tell you more than social media. Single runs vary.
  • Start a fresh session before judging. Very long sessions mix in compaction and stale context.
  • Report with specifics. Dated examples and request IDs, like those in issue #105112, are what a lab can check.
  • Have a fallback. If your work depends on one hosted model, keep a second assistant ready. Our ChatGPT vs Claude vs Gemini vs Grok comparison covers the main options, and subscription vs API costs explains when paying per token gives you more control.

FAQ

Did Anthropic nerf Opus 5.5 on November 3?

There is no evidence of that. The November 3 chart measures the visible thinking summary, which Anthropic's docs say comes from a separate model. One developer's logs show the new summary style arriving on different accounts on different days while reasoning volume held steady.

Is Claude getting worse in general?

Not on fixed test questions: the LiveNerf tracker shows Opus 5.5 at launch level through November 8. Whether it got worse at long, file heavy sessions after October 19 is unresolved, because only Anthropic's logs can confirm or rule out the change reported in issue #105112.

Why does Opus 5.5 feel lazier than Opus 5?

Part of the answer is documented. Opus 5.5 defaults to medium effort while Opus 5 defaulted to high, so a request that sets nothing runs one level lower. Setting effort to high or xhigh is the first thing to try.

Has Anthropic responded to the Opus 5.5 complaints?

As of November 9, 2026, we found no public statement, and the main GitHub issue had no reply from Anthropic. In April 2026 it did publish a postmortem on earlier Claude Code complaints and said it never intentionally degrades its models.

Does Anthropic secretly switch Opus 5.5 to another model?

Only in documented cases. Requests flagged by its cybersecurity or biology safeguards fall back to Opus 4.8 or Opus 5, per the launch post. There is no public evidence of hidden swaps for ordinary requests.