Anthropic has resolved three recent issues that degraded the quality of Claude Code, the Claude Agent SDK, and Claude Cowork. As of April 20, the affected software has been patched, restoring functionality across Sonnet 4.6, Opus 4.6, and Opus 4.7 models. The core API and inference layers remained unaffected throughout the incident. The first issue stemmed from a change on March 4, where the default reasoning effort for Claude Code was switched from high to medium to reduce latency and prevent UI freezing. While this decision aimed to optimize performance for the majority of tasks, users reported a noticeable drop in intelligence. After analyzing feedback indicating a preference for higher quality over reduced latency, Anthropic reverted this change on April 7, restoring high effort as the default for most models. The second problem involved a caching optimization deployed on March 26 designed to clear old thinking data from idle sessions to save costs and improve speed. A critical bug caused the system to delete reasoning history on every turn after a session exceeded an hour of inactivity, rather than just once. This led to models appearing forgetful, repetitive, and prone to poor tool choices because they lost context of their previous steps. The issue also contributed to faster-than-expected usage limit depletion due to increased cache misses. This bug was identified and fixed on April 10. The third incident occurred on April 16 when a new system prompt instruction was added to limit verbosity for the Opus 4.7 model. The instruction aimed to keep text between tool calls under 25 words and final responses under 100 words. However, this constraint inadvertently reduced coding quality and overall intelligence. After additional evaluation revealed a performance drop, the modification was reverted on April 20. These overlapping issues, occurring at different times and affecting different user segments, initially appeared as broad and inconsistent degradation. It took over a week to isolate the root causes, as the problems were difficult to reproduce in internal tests due to specific software environments and experimental features that masked the bugs. In response to these setbacks, Anthropic is resetting usage limits for all subscribers effective April 23. The company has also announced several procedural changes to prevent future occurrences. These include increasing the proportion of internal staff using the exact public build for testing, enhancing the internal Code Review tool, and implementing stricter controls on system prompt modifications. Moving forward, all prompt changes will undergo broader ablation testing and gradual rollouts to ensure stability before reaching users. Anthropic thanked users who reported specific issues and provided reproducible examples via feedback commands or public posts, noting that their input was essential to identifying and resolving these complex bugs.