Skip to main content
Back to timeline
AnthropicSource publication:

Anthropic traces Claude Code quality complaints to three separate changes, all reverted by April 20

Synopsis

Anthropic says a month of user reports that Claude's answers had worsened traces to three separate changes across Claude Code, the Claude Agent SDK and Claude Cowork, while the API was not affected. The changes were a March 4 switch of the default reasoning effort from high to medium, a March 26 attempt to clear old thinking from idle sessions that a bug made fire every turn, and an April 16 system-prompt length limit added to cut verbosity. They were reverted on April 7, April 10 and April 20, and all three issues were resolved as of April 20 (v2.1.116).

Source-provided article image: An update on Claude Code quality reports \ Anthropic
Anthropic

Interpretation

The perceived broad degradation was not one model getting worse but three changes hitting different slices of traffic on different schedules, so the aggregate looked wide and inconsistent. Users had reported that Claude's responses in Claude Code worsened, while Anthropic's internal usage data and evaluations initially did not reproduce the issues. Anthropic confirmed the API and inference layer were unaffected and tied each change to specific symptoms on Sonnet 4.6, Opus 4.6 and Opus 4.7.

The first change lowered Claude Code's default reasoning effort from high to medium to ease occasional very long thinking in high mode that made the UI appear frozen. In internal evaluations medium gave slightly lower intelligence with significantly less latency on most tasks and avoided the long tail of thinking latency, so it became the default. Users said they preferred higher intelligence by default and lower effort for simple tasks, so the change was reverted on April 7; Opus 4.7 now defaults to xhigh and other models to high.

The second change was meant to clear old thinking from sessions idle for over an hour to cut the cost of resuming them, but a bug made it clear on every turn for the rest of the session. Instead of clearing once, each request kept only the most recent block of reasoning and discarded everything before it, including reasoning from the current turn. It surfaced as forgetfulness, repetition and odd tool choices and also caused cache misses and faster usage-limit drain; the bug was fixed on April 10 in v2.1.101.

The third change added a system-prompt length limit to reduce verbosity, and combined with other prompt changes it hurt coding quality. The limit shipped with Opus 4.7 on April 16 only after weeks of internal testing with no regressions in the evaluations run. Ablations over a broader evaluation set during the investigation showed a 3% drop for both Opus 4.6 and 4.7, and the prompt was reverted in the April 20 release.

Perspective

The account covers three changes across Claude Code, the Claude Agent SDK and Claude Cowork, and states the API and inference layer were not affected; the affected models are Sonnet 4.6, Opus 4.6 and Opus 4.7, with fixes in place as of v2.1.116 on April 20.

Anthropic says the reports were at first hard to distinguish from normal variation in user feedback, and internal usage and evaluations initially did not reproduce them; reproducing the caching bug was also obscured by an internal-only server-side message-queuing experiment and a separate change in how thinking is displayed, and it occurred only in the corner case of stale sessions.