Anthropic released a postmortem on the degradations of Claude Code

Anthropic released a postmortem on the degradations of Claude Code. The company acknowledged three issues since early spring that affected code quality. The first was a non-obvious change in the default reasoning effort from high to medium, which lasted more than a month. The second bug removed sessions that had been idle for more than an hour of model reasoning after each message, which degraded quality (fixed in two weeks). The third problem was simply in the system prompt — the model was just asked to generate fewer tokens, which caused it to slow down. Along with the release of the postmortem, all users of Claude Code had their limits reset (the postmortem was released yesterday). In addition, Anthropic promised to use public versions of Claude Code more frequently within the company and to test all prompt changes.