Anthropic attributes Claude Code performance decline to engineering missteps after weeks of user backlash

Anthropic published a postmortem on April 23 attributing weeks of degraded Claude Code output to three engineering missteps, per Fortune: a March 4 change cutting default reasoning effort from high to medium to reduce latency, a March 26 bug that caused the model to discard its own reasoning history mid-session, and an April 16 system prompt capping responses at 25 words between tool calls that was reverted after four days. Anthropic said demand had grown at an unprecedented rate and stretched its infrastructure, particularly at peak hours.

Fortune reports the explanation has not mollified users, some of whom canceled subscriptions and accused the company of initially denying the problems; the piece also cites code-quality testing that found vulnerabilities introduced in 52% of tested tasks versus 30% for OpenAI. The episode underscores how silent serving-stack changes now carry direct churn risk for paid coding-agent products.