Is Claude Opus 5.5 nerfed? The November 3 change was the thinking summarizer; the October 19 step is still unexplained

The chart that has been passed around as proof that Anthropic "nerfed" Claude Opus 5.5 on November 3 shows something real, but probably not what most people think. Developer Shane Melaugh had Gemini 4 Argon go through 1,400 of his Claude Code sessions and found that on November 3, around 18:00 UTC, Opus 5.5's visible thinking suddenly started reading differently. The model ID and the Claude Code version stayed the same. His post, "is Opus 5.5 nerfed?", has been seen more than 440,000 times.

What changed, as far as the available evidence shows, is the text Claude Code displays, not the model producing the answers. Anthropic's documentation says the thinking you see is a summary written by a separate model. Theo Browne's logs suggest that summarizer was swapped account by account over about a week. That leaves a second, quieter complaint unanswered: a step change in answer quality on October 19, measured carefully by one Claude Code user and still without a reply from Anthropic.

What the November 3 chart measures

Melaugh's first chart tracks the share of visible thinking blocks with particular wording. Blocks starting with "I'm…" jumped from 20–30% to 61.3%. Those using "I should" or "I need to" fell from about 50% to 10.9%, and blocks containing a dash fell from 40–55% to 6.5%. The chart says the shift "appeared mid-session across 7 active sessions," and its footnote adds that visible summaries "show a behavior shift, not proof of changed model weights."

That footnote matters, because on Opus 5.5 nobody sees the model's thinking. Anthropic's thinking documentation says the text in a thinking block "is never the raw chain of thought." With display: "summarized" you get "a summary of Claude's full thinking process," and "summarization is processed by a different model from the one you target in your requests. The thinking model does not receive the summarized output." On the API, Opus 5.5 returns no thinking text at all unless the client asks for it, because its default display is "omitted". Billing is for the full thinking tokens, not the summary.

So a change in phrasing like "I'm…" over "I need to" is first of all a change in the summarizer's writing style. It can't tell you whether Opus thinks less or worse. A switch that happens in the middle of a session, with the same client running, also points to the server side.

Theo's account-by-account data

On November 8 Theo Browne (T3) posted his own analysis: 458,000 Opus 5.5 responses that went through his proxy between October 9 and November 7, all on Claude Code 2.1.280 with thinking display set to "summarized." The style change is there too, with "I'm…" openers going from 27% to 61%, "I should/I need to" from 47% to 13%, and the median summary growing from 325 to 440 characters. But it didn't land on one date. His six accounts switched at different times: one on October 30 at 00:00 UTC, one on October 31 at 17:00, one on November 3 at 18:00 (the same hour as the viral chart), and two on November 4. A sixth hadn't switched when he last used it, on November 3.

The measures of the work itself barely moved. Reasoning tokens per response averaged 369 before his accounts switched and 413 after. The reasoning share of output tokens went from 46% to 49%, output speed from 111 to 112 tokens per second, and tool calls per response from 1.03 to 1.06. "Anthropic appears to be rolling out a change to the reasoning summarizer on a per-account basis," he wrote. "No sign of degraded capability for me." Two days earlier he had complained that Opus 5.5's traces "read like a shitty quantized open weight model" even when the answer came out right, which is consistent with the summarizer being the weak part.

Theo's numbers come from his own traffic and haven't been published as raw data, so they're one well-instrumented user's measurements. Still, a rollout spread over six days and several accounts is hard to square with a single model swap on November 3.

We checked the Claude Code release times

One reply to Melaugh offered another explanation: Claude Code releases on November 2 and 3 changed effort handling and dropped old thinking when resuming conversations. That's documented in the CHANGELOG. Version 2.1.317 lets subagents on a different model use that model's own effort level and fixes a project's effort setting being ignored in some desktop sessions. Version 2.1.318 removes an earlier fix that kept a model's earlier thinking when resuming a conversation.

The timing doesn't fit. According to the npm registry, 2.1.317 was published on November 2 at 23:57 UTC and 2.1.318 on November 3 at 21:55 UTC. Melaugh's shift is at about 18:00 UTC, 18 hours after the first release and four hours before the second, and his chart says the CLI version didn't change. Theo was on 2.1.280 throughout. Neither release mentions thinking summaries. The only Claude Platform release note dated November 3 is about text watermarking on three other models (Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6), not Opus 5.5. We found no Anthropic announcement of a summarizer change at all.

What isn't explained by the summarizer

Melaugh's second chart is harder to explain away. For effort set to "high," he reports median thinking tokens before the first tool call falling from 84 to 54 (−36%), an index of "script & edit crashes" rising 47%, and 40.5% of turns emitting no thinking tokens, comparing 14,560 turns before November 3 with 14,871 after. If those are billed thinking tokens, the summarizer can't explain them. If they were estimated from visible summary text, it can. The posts don't say which, and they don't include his method or data. In a separate anecdote, he says the same real bug took Opus 5.5 at high effort three hours without a fix, while xhigh found one in a few minutes.

Melaugh's and Theo's data cover different accounts and workloads, so both could be accurate. But only Theo's includes reasoning-token counts per response, and they rose.

The older step: October 19

The more detailed case for a real regression predates the viral chart. In issue #105112 on the Claude Code repo, GitHub user kjzj9xc5nh-gif scored 6,638 Opus 5.5 answers from local transcripts. A script flagged answers that the user's next message corrected, or that Claude's next reply retracted. 5.2% of 5,738 answers were corrected in the four weeks after launch, compared with 25.4% of 900 afterwards. The user dates the step to between October 18, 8:52 PM and October 19, 1:13 PM Pacific, with the client build (2.1.303), CLAUDE.md, hooks and cached prompt prefix all unchanged.

The failure the user describes isn't a lazy model. Tool calls per answer held steady, and billed thinking rose at medium, high and xhigh effort. The model "reads my files and runs the lookups... and then writes an answer that ignores or contradicts what it read." Raising effort didn't help: at high effort the corrected share went from 5.4% to 26.6%. A November 4 update puts the weekly rate at 23.0%, 20.5% and then 17.0% for the three weeks since, against 3.6–7.0% from late August to mid-October. A second user filed a similar report (#105477) from Windows 11 on Claude Code 2.1.293. As of November 9, neither issue has a reply from Anthropic. Every comment on #105112 comes from users.

It's still one person's workflow, and a "correction" detected by a script is a proxy for quality, not a measure of it. But it's a narrow claim with a narrow time window, and only Anthropic's serving logs can confirm or rule it out.

What the fixed-panel tracker shows across both dates

LiveNerf, an open-source tracker, runs the same 78 questions (GPQA Diamond, MMLU-Pro, competition math, AIME) through headless Claude Code on a Max subscription every day. The CLI is pinned to 2.1.280, and the questions were chosen because Opus 5.5 gets them right only some of the time. Its November 8 README reports no change across three full 10-day windows (window 3: +0.028 ± 0.023 against launch week). The tracker can only detect shifts of about 6.6 points per window.

We pooled the published daily results around both disputed dates:

Nothing moved on either date. Daily median output in the last period (541 to 627 tokens) stays inside the range of the earlier days (493 to 726). This is what a fixed panel can tell you: on single-turn questions with known answers, Opus 5.5 in Claude Code answers as it did at launch.

It can't tell you about the failure in #105112, which shows up in long sessions where the answer depends on files and tool results already in context. LiveNerf doesn't test that, and its README doesn't claim to. The October 19 report isn't refuted by these numbers. It's about something the trackers don't measure.

Where this leaves the "nerf"

The viral November 3 evidence, the wording of visible thinking, is the weakest kind of evidence available for a capability change. Anthropic's docs say that text comes from a separate summarizer, and Theo's data shows that summarizer changing account by account while reasoning volume held steady. A fixed-question tracker sees no change on either side of that date. The stronger evidence is the October 19 step in #105112, which is specific, dated and reproduced in a second report, and which nobody has answered.

Anthropic has been here before. In its April 23 postmortem it traced weeks of Claude Code quality complaints to three product changes: a default effort lowered from high to medium, a bug that cleared Claude's older thinking on every turn instead of once, and a system-prompt line meant to cut verbosity. It said the API and inference layer weren't affected. Opus 5.5 already starts lower: Anthropic's docs say it defaults to medium effort, where most other models that support effort default to high. That's one reason "use high or xhigh" advice keeps coming up. What would settle the question is short: confirmation that the thinking summarizer changed and when, and an answer to whether anything about how claude-opus-5-5 is served changed between October 18 and 19.