A noticeable deterioration in the performance of the Claude artificial intelligence model has been recorded. AMD's Head of AI analyzed Claude's sessions from January to March, revealing an unpleasant picture.
The editorial office reports that the model's reasoning ability has decreased. The average length of deliberations has dropped from approximately 2,200 characters to 600 characters, which immediately impacts the quality of decisions. At the same time, the number of API requests from February to March has increased by almost 80 times; less analysis, more attempts, and higher token consumption are observed.
The model's behavior has also changed. Claude more frequently "gives up" or asks whether to continue working; while such instances were not observed before March 8, 173 such incidents were recorded within 17 days. Also, the number of views per edit (the number of file/code sections the model reviewed before making a change) has dropped from 6.6 to 2.0, indicating that Claude examines code less thoroughly before making changes.
Furthermore, the number of contradictions has increased, and the model more frequently attempts to relearn during the response process. Users note that Claude has started to disregard information like CLAUDE.md, as it lacks sufficient "thinking budget" to consider the context. Interestingly, quality depends on the time of day: the worst results are recorded between 5:00 PM and 7:00 PM PST, while the best results are observed at night, which appears to be directly related to GPU load.
AMD's Director of AI analyzed 6,852 sessions of Claude Code, revealing a significant deterioration in the model. This analysis covered 234,760 tool calls, 17,871 reasoning blocks, and 3 months of logs. Anthropic responded to these findings, essentially confirming these conclusions.