Claude Opus 5.5 Quietly Changed on Nov 3
Original: Is Opus 5.5 nerfed? Thinking patterns changed on November 3
Why This Matters
Silent model updates undermine developer trust and reproducibility at scale.
Developer Shane Melaugh analyzed 1,400 Claude Code sessions with Gemini 4 Argon and found Opus 5.5's reasoning patterns shifted sharply on November 3, with no model name change. The 'high' thinking tier saw a notable drop in reasoning depth.
Shane Melaugh, builder of SashimiiAI, noticed a sudden drop in Opus 5.5 performance and ran 1,400 Claude Code sessions through Gemini 4 Argon to investigate. The analysis pointed to November 3 as the inflection point—model name unchanged, but reasoning behavior clearly altered. Specifically, the amount of thinking Opus 5.5 performs at the 'high' budget level dropped sharply. In a practical test on a real bug, 'high' mode jumped too quickly into action, spent three hours failing to find a fix. The same task on 'xhigh' resolved in minutes. Melaugh's conclusion: the widely-cited advice from creator Theo to 'use high for everything' no longer holds. He added a caveat that his analysis isn't a formal benchmark—it started from personal frustration after his workflow degraded. The post drew 531K views, with Theo himself spinning up a job to investigate his own logs via ClickHouse. Anthropic has not publicly acknowledged any change.