Anthropic revealed that a combination of prompt changes, caching bugs, and reasoning effort tweaks caused performance degradation across Claude Code, Claude Agent SDK, and Claude Cowork. A restrictive verbosity prompt alone triggered a 3% drop in evaluation performance across Opus 4.6 and Opus 4.7 before being reverted in version 2.1.116. To compensate affected developers, the company reset usage limits for all subscribers following its postmortem investigation.
Three Root Causes Behind the Coding Quality Drop
According to an engineering postmortem published by Anthropic, the degradation observed by users stemmed from three distinct changes rolled out between March and April. The underlying API inference layer remained unaffected, but client-side software harnesses suffered cumulative quality loss.
First, Anthropic changed Claude Code's default reasoning effort from high to medium on March 4 to address long latency spikes. After user feedback indicated a clear preference for intelligence over speed, this change was reverted on April 7, setting defaults to xhigh for Opus 4.7 and high for other models.
Second, an April 16 system prompt addition aimed at reducing verbosity instructed the model to keep text between tool calls under 25 words. Internal ablations later confirmed this strict limit severely impacted coding reasoning, prompting an immediate revert on April 20.
Caching Bugs and Context Loss
The third and most complex issue involved a caching optimization introduced on March 26. Intended to clear older thinking sections from sessions idle for over an hour, a software bug caused the system to drop reasoning history on every subsequent turn using the clear_thinking_20251015 API header.
Key timeline milestones from Anthropic's postmortem include:
- March 4 – April 7: Reduced default reasoning effort created perception of lower intelligence in Sonnet 4.6 and Opus 4.6.
- March 26 – April 10: Logic bug continuously wiped reasoning blocks, causing cache misses and rapid usage limit exhaustion.
- April 16 – April 20: Verbosity prompt restrictions triggered a 3% drop in benchmark evaluations across multiple models.
Because reasoning history was repeatedly stripped, subsequent requests resulted in frequent cache misses, driving up token costs and draining subscriber usage caps faster than expected.
Detection Challenges and Future Safeguards
Anthropic acknowledged an honest caveat in its quality assurance pipeline: internal unit tests, dogfooding, and end-to-end verifications failed to catch the caching bug initially. Internal-only server experiments and differences in CLI thinking displays masked the issue during internal testing for over a week.