Update on Recent Claude Code Quality Reports

Anthropic said it investigated recent reports from some users that Claude response quality had gotten worse, and confirmed the cause was 3 separate changes spanning Claude Code, the Claude Agent SDK, and Claude Cowork. The API and inference infrastructure were unaffected, and all three issues were fixed as of April 20 (v2.1.116).

The main causes are as follows.

• Default reasoning effort change: On March 4, Claude Code's default was lowered from high to medium in an attempt to reduce latency issues, but this led to quality degradation, so it was reverted on April 7. This impact spanned Sonnet 4.6, Opus 4.6.
• Session thinking cache removal bug: On March 26, an optimization was added to clear stale thinking from sessions that had been idle for over an hour, but due to a bug, thinking continued to be cleared every turn throughout subsequent sessions. As a result, it appeared as memory loss, repetition, and odd tool choices, and was fixed on April 10. This also affected Sonnet 4.6, Opus 4.6.
• Verbosity-reduction system prompt change: A prompt added on April 16 to encourage shorter answers combined with other changes to degrade coding quality, and was reverted on April 20. This impact spanned Sonnet 4.6, Opus 4.6, Opus 4.7.

Anthropic explained that these three changes affected different traffic at different times, which together appeared as broad but uneven performance degradation overall. It also noted that these were difficult to reproduce through internal usage and initial evals, and in particular the second bug passed multiple reviews, tests, and dogfooding, which ultimately led to pursuing an improvement to include more repository context in code review.

Going forward, Anthropic said it would strengthen the following.

• Expanding internal staff use of public builds
• Improving the internal Code Review tool and deploying it to customers
• Applying broader model-specific evals, ablations, soak periods, and gradual rollouts to system prompt changes
• Strengthening CLAUDE.md guidance so that model-specific changes are limited to that model only

Finally, Anthropic said that users reporting specific cases via /feedback was what enabled it to find and fix the issue, and added that it has reset usage limits for all subscribers.