Anthropic Reveals Causes of Claude Performance Degradation

Anthropic investigated reports of degraded Claude response quality raised over the past month and found that 3 separate changes were responsible, each affecting Claude Code, Claude Agent SDK, and Claude Cowork respectively. It explained that the API and inference layer were not affected.

The main causes are as follows.

• Change to Claude Code's default reasoning effort: On March 4, the default was lowered from high to medium, but Anthropic later decided it was better to default to higher intelligence and let users opt into lower effort, so it reverted the change on April 7. This change affected Sonnet 4.6, Opus 4.6.
• Bug deleting thinking in idle sessions: On March 26, Anthropic tried to reduce latency by clearing old thinking in sessions that had been idle for more than 1 hour, but a bug caused thinking to keep being deleted on every turn in the session from then on. As a result, Claude showed repetitive and strange tool choices as if it had lost its memory, and cache misses also caused usage limits to be consumed faster. This issue was fixed on April 10.
• System prompt change to reduce verbosity: On April 16, a prompt to shorten responses was added, but combined with other prompt changes it degraded coding quality, so it was reverted on April 20. It affected Sonnet 4.6, Opus 4.6, Opus 4.7.

Anthropic explained that because these three changes occurred at different times and across different traffic, they appeared externally as irregular, overall quality degradation. It added that since internal usage and evals alone did not immediately reproduce the issue, it took time to identify the causes.

To prevent recurrence, Anthropic said it will introduce expanded use of public builds, improved internal Code Review and deployment to customers, stricter controls over system prompt changes, broad evals and ablations per model, and soak periods and gradual rollout. It also announced that, as of April 23, 2026, it will reset usage limits for all subscribers.