Benchmarks say Opus 5.5 is fine. Devs say it ate a crayon.

• AI Stupid Level (opens in a new tab) (as of Nov 7, 2026): Combined score about 94 ± 0.3, trend holding steady; "no sustained decline detected" on coding, reasoning, or tool use.
• LiveNerf (opens in a new tab) (day 45, Nov 7): three completed 10-day windows after launch week, no change detected on its frozen panel (window 3: +0.028 ± 0.023 vs launch week).
• BridgeBench NerfBench (opens in a new tab) (Oct 16 weekly post): Opus 5.5 at 105.8% of launch; every run cited in coverage stayed inside the 90–110% band; headline "No nerfs detected."

The strongest complaint log. In anthropics/claude-code#105112 (opens in a new tab), one power user scored their own Claude Code transcripts: corrected answers 5.2% of 5,738 before Oct 19 vs 25.4% of 900 after. Those are that author's measurements, not an Anthropic figure and not independently verified by MODEL. Their claim is not "lazy": tool calls and billed thinking held or rose; the failure mode is research done, then an answer that ignores what it just read.

A documented confounder (not a mid-life nerf). Anthropic's effort docs (opens in a new tab) say Claude Opus 5.5 and Claude Haiku 5.5 default to medium; earlier Opus models defaulted to high, so a request that omits effort runs one level lower than it did on Opus 5. That can make the model feel lighter without any silent weight change.

Anthropic has owned harness quality bugs before. In its April 23, 2026 Claude Code quality postmortem (opens in a new tab), the company traced weeks of "it got worse" reports to three product changes — a default effort cut, a caching bug that cleared prior thinking, and a verbosity prompt — not to intentional model degradation. All three were fixed by April 20 (v2.1.116). That history is context, not proof of a 2026-10 Opus 5.5 nerf.

Caveat (both can be true). Trackers mostly measure fixed or short API-style tasks. The loudest complaints are long, stateful agent sessions. A model can hold a frozen panel and still loop a rejected fix nine times in Claude Code. Do not read this as Anthropic confirming a nerf — they have not, publicly, as of Nov 7.

Disclosure: Not sponsored. No affiliate links. Not tested by MODEL. GitHub correction rates are the issue author's own logs.

MODEL Editors