AI News · 6 min read
Is Claude Opus 5.5 nerfed? What the November 3 data shows
A 1,400-session analysis says Opus 5.5's reasoning changed on Nov 3. Theo's 458k-response logs say only the summaries did. What's verified and what to do.
What set it off
On November 7, 2026 at 22:15 UTC, developer Shane Melaugh asked on X ↗: "is Opus 5.5 nerfed?" He said he'd had Gemini 4 Argon analyze 1,400 of his Claude Code sessions, and 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 passed 450,000 views. On Monday a Hacker News thread ↗ titled "Is Opus 5.5 nerfed? Thinking patterns changed on November 3" drew 137 points and 78 comments within about three hours.
Anthropic hadn't commented publicly as of Monday morning, November 9 (UTC). There's no status-page incident or changelog entry that mentions a model change. So for now, this is a debate between two sets of user logs.
The case for a nerf
Melaugh's first chart tracks the visible thinking blocks in his sessions. Around 18:00 UTC on November 3, the share starting with "I'm…" jumped to about 61%. Blocks using "I should / I need to" fell to about 11%, and blocks with dashes fell to about 7%. The chart says the change "appeared mid-session across 7 active sessions," with the same model ID and CLI version.
His second chart ↗ is the one that worries people. He compared 14,560 turns at high effort from before November 3 with 14,871 from after. Median thinking tokens before the first tool call fell from 84 to 54 (−36%), and an index of script and edit crashes rose 47%. In a follow-up ↗, he described a real bug that Opus 5.5 on high couldn't fix in "3 full hours," while xhigh "finds a fix after a few minutes." His conclusion: "the nerf is real," and the popular advice to "use high for everything" no longer applies.
Plenty of people on Hacker News agreed. "Opus is suddenly fumbling basic tool calls and failing to finish tasks even at high," wrote one commenter. Others reported faster usage burn. To be fair, Melaugh's own chart carries the caveat: "Visible summaries show a behavior shift, not proof of changed model weights."
The case against
Theo Browne (t3.gg) pointed an agent at his own proxy logs. His conclusion ↗: "Anthropic appears to be rolling out a change to the reasoning summarizer on a per-account basis. No sign of degraded capability for me."
His chart covers 458,000 Opus 5.5 responses from October 9 to November 7, and it explains the timing confusion. Each of his accounts switched to the new summary style at a different hour, between October 30 and November 4. One of them switched at 18:00 UTC on November 3, the same hour as Melaugh's chart. The summary text changed the same way: "I'm…" openings went from 27% to 61%, and "I should / I need to" fell from 47% to 13%. The work didn't change:
| Theo's logs | Before switch | After switch |
|---|---|---|
| Reasoning tokens per response | 369 | 413 |
| Reasoning share of output | 46% | 49% |
| Output speed (tok/s) | 111 | 112 |
| Tool calls per response | 1.03 | 1.06 |
Another developer, @argofowl ↗, checked 67,000 of their own Opus messages. Tool calls per message stayed at 1.06, and failed tool calls went from 1.9% to 1.8%. "Not seeing a nerf on my account," they wrote.
What Anthropic's docs and changelog actually say
Two verifiable facts matter here.
1. The thinking you see is not the model's thinking. Anthropic's thinking documentation ↗ says that with summarized display, "the thinking text you receive is a summary of Claude's full thinking process rather than the raw chain of thought," and that "summarization is processed by a different model from the one you target." You're billed for the full thinking tokens, not the summary. So a change in summary wording is exactly what a summarizer change would look like. It says nothing either way about Opus itself.
2. Claude Code did change that week, but not in a way that explains this. Per the Claude Code changelog ↗, version 2.1.317 (published to npm November 2, 23:57 UTC) changed subagents on a different model to use that model's own saved effort level instead of the session's. Version 2.1.318 (November 3, 21:55 UTC) removed an older fix that preserved earlier thinking when resuming some conversations. Neither mentions the summarizer. 2.1.318 also landed hours after the 18:00 UTC shift, and both Melaugh and Theo say their CLI version didn't change.
That leaves Melaugh's thinking-token drop as the strongest nerf evidence. It's also unexplained: Theo's average reasoning tokens went up. These are different measurements (median tokens before the first tool call vs. average per response) on different workloads, so they can both be true. Neither proves Anthropic changed the model.
Our take
Based on public evidence, the "nerf" is not proven, and the summary change is close to proven. Theo's per-account switch times line up with Melaugh's shift down to the hour, and Anthropic's docs say a separate model writes those summaries. If you've been judging Opus by how its thinking reads, you've been grading the narrator.
That doesn't mean the complaints are imaginary. Agent results vary a lot by task, and a 36% drop in upfront planning at high effort is worth taking seriously if it shows up in your own logs. Here's what we'd do:
Measure outcomes, not vibes. Claude Code keeps session transcripts on disk (as
.jsonlfiles under~/.claude/projects). Compare test pass rates, retries and failed tool calls before and after November 3 on your repo.Try xhigh on hard bugs. Several people, including Melaugh, report better results there. It costs more, so test it on the tasks where high fails.
Keep a second model handy. Opus 5.5 is still at or near the top for agentic coding in our best AI for coding ranking, but pinning your whole workflow to one vendor's silent rollouts is the real risk. Our Codex vs Claude Code comparison and Claude Code review cover the fallbacks.
The fix that would end this argument is simple: Anthropic should log summarizer changes in its release notes. If a separate model rewrites what users see, users deserve to know when it changes. We'll update this post if Anthropic responds.
Sources
Shane Melaugh on X ↗ and follow-ups ↗, November 7, 2026
Theo (@theo) on X ↗, November 8, 2026
@argofowl on X ↗, November 8, 2026
Hacker News: "Is Opus 5.5 nerfed?" ↗, November 9, 2026
Anthropic: Thinking documentation, "Summarized thinking" ↗, checked November 9, 2026
Claude Code CHANGELOG ↗ (2.1.317 and 2.1.318) and npm publish times, checked November 9, 2026
Reddit discussion was linked from Hacker News but wasn't reachable from our network (HTTP 403), so reactions here come from X and HN only.