Claude Code got dumber on purpose, then by accident, then on a word count
Ourtake
Three separate tweaks, one vague sense that the assistant had gotten worse. Anthropic says the eval suite will catch the next one.
What happened
Anthropic admitted that three changes to Claude Code degraded its quality: a lowered default reasoning effort from March 4, a caching bug shipped March 26 that wiped reasoning history each turn, and an April 16 system prompt length limit that caused a 3 percent quality drop. Anthropic rolled back the changes between April 7 and April 20 and promised stricter testing and rollouts.