Anthropic's Claude Code Incident Postmortem: April 23, 2025

Anthropic has published an official postmortem explaining the causes of the Claude response quality degradation reported over the past month. The affected scope was Claude Code, Claude Agent SDK, and Claude Cowork, and the API itself was not affected. All issues were resolved as of April 20, 2025 (v2.1.116).

Three changes caused problems in different ways.

• March 4: Claude Code's default reasoning effort was lowered from high to medium, which reduced slow responses but was confirmed to cause quality degradation, so it was reverted on April 7.
• March 26: A caching optimization that cleared previous thinking records when resuming sessions idle for more than 1 hour had a bug that caused it to run repeatedly on every subsequent turn. This caused Claude to lose past context, resulting in repetitive responses, abnormal tool selection, frequent cache miss, and faster-than-expected usage consumption. This issue was fixed on April 10.
• April 16: A system prompt was added aiming to reduce verbosity in Opus 4.7, limiting text between tool calls to 25 words or less and final responses to 100 words or less. This wasn't revealed in internal testing but degraded actual coding quality in practice, so it was removed on April 20.

Anthropic also disclosed why the issue took so long to discover. The changes were applied at different times and to different traffic scopes, making the causes appear scattered, and differences between internal testing environments and real user environments made reproduction difficult. The existing eval suite also wasn't broad enough, and the impact of the system prompt change was only confirmed as a 3% performance drop after running broader evaluations.

Preventive measures going forward include:

• Mandatory use of public builds by internal employees
• Per-model evaluations, ablations, gradual rollout, and sufficient soak periods for system prompt changes
• Expanded repository scope that code review can reference
• A new user communication channel, @ClaudeDevs

Anthropic stated that it never intentionally degraded the model, and it has reset usage limits for subscribers.