How Anthropic Nerfed Opus 4.6 Before the 4.7 Launch

Anthropic released Claude Opus 4.6 on Feb 5, 2026. On Feb 9, adaptive thinking became default, letting the model choose its reasoning budget. On Mar 3, default effort dropped from high to medium. Between Mar 5-12, a beta header hid reasoning from the UI. By Mar 12, all reasoning was hidden. These were product changes, not weight updates. Stella Laurenzo of AMD analyzed 6,852 sessions. Median visible reasoning per turn fell 73% (from 2,200 to 600 characters). Read-to-edit ratio dropped from 6.6 to 2.0. User interruption rate rose twelvefold, from 0.9 to 11.4 per 1,000 tool calls. Edits to files not recently read jumped from 6.2% to 33.7%. She filed GitHub issue #42796 on Apr 2. Claude Code lead Boris Cherny confirmed the defaults and said adaptive thinking sometimes allocates zero reasoning per turn, causing fabrication. Workaround: CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1. On Apr 16, Anthropic released Opus 4.7 with default effort reset to 'xhigh,' higher than both previous defaults. Users saw improvement. A separate viral claim that Opus 4.6 dropped from rank #2 to #10 was debunked: it compared different test sets. The version string and price didn't change between Feb and Apr, but the product degraded temporarily.