Anthropic has admitted to secretly degrading the reasoning capabilities of its Claude AI model for six weeks. The issue began on March 4th, when the model's reasoning effort was reduced from HIGH to MEDIUM, leading to shorter chains of thought and increased errors on complex tasks. This degradation was compounded by a caching bug on March 26th that deleted reasoning history, causing benchmark accuracy to drop by 18 points. Anthropic only disclosed these issues on May 28th, coinciding with the launch of Opus 4.8. AI
IMPACT This incident highlights the importance of transparency in AI model performance and the potential for undisclosed degradations to impact users.
RANK_REASON The cluster discusses a past event and Anthropic's admission, rather than a new release or development.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
