Anthropic fixes Claude Code bugs
- Anthropic said on April 23 that three separate Claude Code changes degraded coding quality for weeks, and all were fixed by April 20. - The postmortem says a March 26 caching bug repeatedly erased prior reasoning after idle sessions, making Claude Code forgetful, repetitive, and costlier to use. - Anthropic reset subscriber usage limits and said its API was unaffected. (anthropic.com)
Anthropic says recent Claude Code quality problems were real, traced to three separate changes, and fixed in version 2.1.116 by April 20. (anthropic.com) Claude Code is Anthropic’s coding agent: it reads a codebase, edits files, runs tests, and returns working changes instead of just suggesting snippets. Anthropic said the problems hit Claude Code, the Claude Agent SDK, and Claude Cowork, but not its API or inference layer. (anthropic.com 1) (anthropic.com 2) The first change landed March 4, when Anthropic lowered Claude Code’s default reasoning effort from high to medium to cut long delays that made the interface look frozen. Anthropic said users preferred the higher-intelligence default, so it reverted that change on April 7. (anthropic.com) The second issue arrived March 26 in a caching change meant to speed up sessions resumed after more than an hour of inactivity. Anthropic said a bug kept clearing older reasoning on every later turn, making Claude seem forgetful and repetitive until the fix on April 10. (anthropic.com) (the-decoder.com) The third problem came on April 16, when Anthropic added a system instruction to make responses shorter. Anthropic said that prompt, combined with other prompt changes, hurt coding quality and was rolled back on April 20. (anthropic.com) Anthropic said each issue hit a different slice of traffic on a different schedule, which made the decline look broad but inconsistent. That pattern helps explain why developers spent weeks describing Claude Code as “dumber” or less reliable without seeing one single failure mode. (anthropic.com) (venturebeat.com) The company said its internal usage and evaluations did not initially reproduce the failures users were reporting in early March. In response, Anthropic said it will tighten pre-release testing and has reset usage limits for all subscribers as of April 23. (anthropic.com) (the-decoder.com) The postmortem is unusually direct for an AI product complaint that had been circulating in public for weeks. Anthropic’s bottom line was simpler: the model was not intentionally degraded, the API was unaffected, and the fixes are now live. (anthropic.com)
