Anthropic Publishes Postmortem After Claude Code Quality Degraded for Weeks Without Detection
Three separate engineering changes between March and April caused Claude Code users to experience forgetfulness, repetition, and reduced coding intelligence — issues that persisted for weeks before all were resolved on April 20.
Anthropic published a detailed postmortem on April 23 disclosing that Claude Code, its AI-powered coding assistant, suffered three overlapping quality regressions between early March and April 20 that went undetected for extended periods. The issues affected users of Claude Sonnet 4.6, Opus 4.6, and Opus 4.7 in the desktop client; the API was not affected.
The first regression, introduced on March 4, changed the default reasoning effort in Claude Code from high to medium in order to reduce latency. The change lowered the model's effective intelligence for coding tasks.
It was not identified until later and was reverted on April 7, restoring reasoning effort to high and extra-high settings. A second bug introduced on March 26 caused a flaw in idle session caching that cleared the model's accumulated reasoning at the start of each turn, producing responses that appeared forgetful or repetitive.
That fix shipped on April 10 as version 2.1.101.
A third problem arose on April 16 when a system prompt was added to reduce verbosity, instructing the model to keep responses under 25 words between tool calls and under 100 words for final answers. The constraint hurt coding quality by preventing the model from adequately planning and explaining its work.
It was reverted on April 20 with the release of version 2.1.116.
Anthropic acknowledged that its evaluation systems did not catch the issues quickly enough and that the combination of three simultaneous but unrelated regressions made the root cause difficult to identify. The company said it has reset usage limits for all affected subscribers and committed to broader pre-deployment evaluations, mandatory audits of system prompt changes, and gradual rollouts for future updates.
Read the original reporting at Anthropic.
