Claude degradation issues have been resolved by Anthropic, which identified and fixed three separate problems that caused Claude’s responses to decline for some users over the past month, the company said in a detailed postmortem published April 23. All three problems have been fixed as of April 20 in version 2.1.116, with the API and inference layer remaining unaffected throughout.
The Claude degradation issues stemmed from changes to Claude Code, the Claude Agent SDK, and Claude Cowork. The first involved a shift in default reasoning effort from high to medium on March 4, which was intended to reduce latency but resulted in lower output quality. Anthropic reverted this change on April 7 after users reported preferring higher intelligence by default, affecting Sonnet 4.6 and Opus 4.6 models.
A second problem emerged from a caching optimization deployed March 26. The change was designed to clear older reasoning from idle sessions lasting over an hour, but a bug caused it to clear reasoning on every turn instead of once. This made Claude appear forgetful and repetitive, as the model lost context of its prior decisions. The issue was particularly compounded when users sent follow-up messages while Claude executed tool commands. Anthropic fixed this on April 10 in version 2.1.101, affecting Sonnet 4.6 and Opus 4.6.
Claude degradation issues traced to system prompt change
The third issue involved a system prompt addition designed to reduce verbosity in Claude Opus 4.7. On April 16, Anthropic added a length-limiting instruction stating: “Keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail.” While internal testing showed no regressions, broader evaluations later revealed a 3 percent performance drop for both Opus 4.6 and 4.7. The company reverted this change on April 20.
Each issue affected different user segments on different schedules, creating the appearance of broad, inconsistent degradation. While Anthropic began investigating reports in early March, distinguishing them from normal variation in user feedback proved challenging at first. Neither internal usage patterns nor evaluation tests initially reproduced the problems.
Reproduction challenges delayed identification
The caching bug proved particularly difficult to identify. Two unrelated experiments complicated reproduction efforts: an internal-only server-side experiment related to message queuing, and an orthogonal change in how thinking was displayed that suppressed the bug in most command-line interface sessions. The bug existed at the intersection of Claude Code’s context management, the Anthropic API, and extended thinking capabilities. It passed multiple human and automated code reviews, unit tests, end-to-end tests, automated verification, and internal dogfooding before discovery.
During investigation, Anthropic used Opus 4.7 to retrospectively review the code repositories that introduced the bug. Opus 4.7 identified the issue while Opus 4.6 did not, prompting the company to commit to supporting additional repositories as context for future code reviews.
Operational changes to prevent recurrence
Anthropic outlined several changes to prevent similar issues. The company will ensure a larger share of internal staff use the exact public build of Claude Code rather than internal testing versions. It will also improve its internal Code Review tool and ship the improved version to customers.
Additionally, Anthropic is implementing tighter controls on system prompt changes. The company will run a broad suite of per-model evaluations for every system prompt modification, continuing ablations to measure each line’s impact. New tooling has been built to simplify prompt change reviews and auditing. The company has updated its CLAUDE.md guidance to ensure model-specific changes apply only to their intended models.
For changes that could trade off against intelligence, Anthropic will add soak periods, broader evaluation suites, and gradual rollouts to catch issues earlier. The company recently created a @ClaudeDevs account on social media to explain product decisions and reasoning in depth, with similar updates shared in centralized threads on GitHub.
Usage limits reset for all subscribers
Anthropic reset usage limits for all subscribers on April 23 as compensation for the degraded experience. The company credited users who provided detailed feedback through the feedback command or posted specific, reproducible examples online for enabling identification and resolution of the problems.
“We take reports about degradation very seriously. We never intentionally degrade our models.”
Anthropic, April 23 Postmortem
The detailed postmortem demonstrates Anthropic’s approach to addressing performance issues in its AI models and commitment to transparency with users regarding technical problems and their resolution.
Source: anthropic.com
