Anthropic Published a Postmortem on Three Claude Code Bugs That Degraded Quality for Weeks

Anthropic Published a Postmortem on Three Claude Code Bugs That Degraded Quality for Weeks

Anthropic published an engineering postmortem on April 23, 2026, disclosing three separate bugs that degraded Claude Code quality between early March and April 20. The bugs affected reasoning effort, context retention, and response verbosity — the core dimensions of Claude Code's usefulness for developers. Anthropic has reset usage limits for all affected subscribers and committed to structural changes in how system prompt changes are tested and reviewed.

Bug One: Reasoning Effort Quietly Downgraded

On March 4, 2026, the default reasoning effort level was changed from "high" to "medium." This was not communicated to users. The result was measurably lower quality on complex reasoning and coding tasks. The change was only reversed on April 7 — more than a month later — after customer complaints escalated. Anthropic has not explained why a month passed before the regression was identified and reversed.

Bug Two: Thinking Cache Cleared Every Turn

From March 26 to April 10, a bug caused the thinking cache to clear on every conversation turn instead of only during idle periods. The practical effect was that Claude Code appeared forgetful — losing context that should have been retained across a session. The bug also caused unexpected usage consumption, as the model was effectively reprocessing context it should have retained. Users who noticed degraded behavior during this window were experiencing a real regression, not imagined issues.

Bug Three: Verbosity System Prompt Throttled Responses

From April 16 to April 20, a system prompt limited Claude Code responses to 25 words or fewer between tool calls and 100 words for standard responses. This constraint degraded coding output quality by approximately 3% on benchmarks — and significantly more in practice for tasks requiring detailed explanations or multi-step code review. The prompt was part of a latency optimization experiment that was not sufficiently tested before deployment.

What Anthropic Is Changing

Anthropic stated it will implement improved testing protocols for system prompt changes and better governance around what changes require explicit user notification. All subscriber usage limits were reset to compensate for the degraded service period. Version 2.1.116 contains all three fixes as of April 20.

My Take

The bugs themselves are understandable — complex systems have regressions. What is harder to excuse is the timeline: the reasoning effort downgrade persisted for 34 days before reversal, and users were left to wonder why their AI coding tool was performing worse. The postmortem is a good step. The question is whether the governance changes are structural or cosmetic. Developers relying on Claude Code for production workflows needed to know about these regressions in real time, not retrospectively.

Related Articles

Sources