Claude Code Quality

After investigating reports of Claude quality degradation over the past month, Anthropic found the cause to be 3 separate changes spanning Claude Code, Claude Agent SDK, and Claude Cowork. The API and inference layer were not affected.

All issues were resolved as of April 20 (v2.1.116).

• March 4: Claude Code's default reasoning effort was lowered from high → medium to reduce long latency, but this turned out to be the wrong trade-off and was reverted on April 7. Affected Sonnet 4.6 / Opus 4.6.
• March 26: An optimization that cleared old thinking from sessions idle for over 1 hour introduced a bug, causing thinking to be continuously deleted every turn across all subsequent sessions. This resulted in memory loss, repetition, odd tool selection, and accelerated usage limit consumption, and was fixed on April 10 (v2.1.101). Affected Sonnet 4.6 / Opus 4.6.
• April 16: A system prompt phrase added to reduce verbosity ended up degrading intelligence and was reverted on April 20. Affected Sonnet 4.6 / Opus 4.6 / Opus 4.7.

To prevent recurrence, Anthropic said it will increase internal use of public builds, strengthen code review tools, and apply broader evals, ablations, soak tests, and gradual rollouts for every system prompt change. On April 23, it also reset usage limits for all subscribers.