Claude Opus 4.6 Silent Degradation Timeline (February-March 2026)
Anthropic silently changed Claude Opus 4.6 in three steps between February and March 2026, reducing reasoning effort and hiding thinking content from users so the regression couldn't be diagnosed.
Between February 9 and March 12, 2026, Anthropic made three undisclosed changes to Claude Opus 4.6 that collectively degraded its performance on complex tasks. On February 9, 2026, the company introduced adaptive thinking, letting the model decide per-turn how many reasoning tokens to spend. Boris Cherny later confirmed that turns with zero reasoning tokens corresponded exactly to the turns where the model fabricated content. On March 3, 2026, the default effort level was silently dropped from high to medium. Boris Cherny defended the change publicly, arguing that 'level 85 is the best balance.' Between March 4 and March 12, 2026, the visibility of thinking content shown to users was redacted from 100% to under 1% over eight days. This made it nearly impossible for users to diagnose why outputs had become shorter, less thorough, and more error-prone, because they could no longer inspect the chain of reasoning. None of these changes were announced. Users experienced a cumulative regression — shorter outputs, more skipped steps, more hallucination — without the diagnostic information needed to file precise bug reports. The pattern was first reconstructed publicly through telemetry analysis by Stella Laurenzo of AMD in early April. See also Claude Code: Clear Context vs Keep Context for Complex Tasks for related context-management behavior.