Developers dispute whether Claude Opus 5.5 has degraded

Developers are publicly questioning whether Anthropic's Claude Opus 5.5 has lost reasoning ability, according to GIGAZINE, with analyses of Claude Code usage logs producing conflicting conclusions.

Shane Melaugh examined about 1,400 Claude Code usage records and said he saw a clear change in the model's thinking-summary sentences and working style around the 3rd of the month. Using Google's Gemini 4 Argon to analyze records from the 9th of last month to the 7th of this month, he found the share of thinking summaries beginning with “I'm” rose from 20-30% to 61.3%, while phrases such as “I should” and “I need to” fell from about 50% to 10.9%. Comparing roughly 29,400 interactions around the 3rd, he said the median number of thinking tokens shown before the first task dropped from 84 to 54, about 36%, while code execution and correction error indicators rose 47%.

In a test to fix the same audio playback bug, Melaugh said the model with reasoning set to “high” failed after 160 interactions, while the “xhigh” setting fixed it in 5.

Second analysis finds no performance drop

Theo Browne analyzed about 458,000 responses and also saw the style of thinking summaries change, but the timing differed by account, and average thinking tokens per response rose from 369 to 413. Browne said he found no sign of performance decline in his environment and raised the possibility that Anthropic is changing its thinking-summary method account by account.

Anthropic's official documentation states that the thinking text shown to users is a summary generated by a separate model, not the AI's actual internal reasoning. Melaugh acknowledged his investigation was not a strict performance evaluation. It has not been confirmed whether Anthropic lowered the model's performance, and further verification is needed.