Anthropic Resolves Claude Code Quality Issues After Month-Long Investigation
Key Takeaways
- ▸Three distinct bugs in Claude Code, the Claude Agent SDK, and Claude Cowork caused quality degradation between March 4 and April 20, though the API remained stable
- ▸Users preferred higher intelligence with longer latency over the reduced reasoning effort Anthropic had switched to, leading to the reversion of the March 4 change
- ▸A session memory clearing bug caused unexpected forgetfulness and repetition due to incorrect implementation logic, fixed on April 10
Summary
Anthropic has identified and resolved three separate issues that caused degraded performance in Claude Code, its AI agent SDK, and Claude Cowork over the past month. The problems stemmed from changes made in early-to-mid April: a March 4 shift of Claude Code's default reasoning effort from high to medium to address latency concerns, a March 26 bug that caused Claude to repeatedly clear session memory instead of once, and an April 16 verbosity reduction that impacted coding quality. All issues have been fixed as of April 20 (version 2.1.116), and Anthropic has reset usage limits for all subscribers as of April 23.
The investigation revealed that each issue affected different user segments on different schedules, creating the appearance of broad, inconsistent degradation that was initially difficult to distinguish from normal variation. While Anthropic began investigating reports in early March, neither internal usage metrics nor their evaluation benchmarks initially reproduced the problems. The company emphasized that the API and inference layer remained unaffected throughout the incident, and that the issues were specific to the Claude Code product interface and related systems.
- Anthropic is implementing process improvements to detect and prevent similar issues faster, and has reset subscriber usage limits as compensation
Editorial Opinion
Anthropic's transparent post-mortem on these quality issues demonstrates both accountability and a user-centric approach to product decisions. However, the month-long gap between initial reports (early March) and full resolution (April 20) raises questions about monitoring and debugging velocity for a company positioning itself as a safety-focused AI leader. The decision to default to lower reasoning effort backfired because it prioritized latency over the core value proposition that users actually wanted—this suggests Anthropic should weight user experience more heavily alongside infrastructure metrics in future optimization decisions.
