Anthropic Resolves Claude Code Quality Regressions
Anthropic identified and fixed three independent regressions that degraded the Claude Code, Claude Agent SDK, and Claude Cowork experiences for some users. The core API and inference layer were not affected. The root causes were a default reasoning effort reduction (high to medium) introduced March 4, a session-state clearing bug shipped March 26 that repeated every turn, and a verbosity-reduction system prompt added April 16. All three issues were reverted or corrected by April 20 in release v2.1.116, and usage limits were reset for subscribers on April 23. The postmortem emphasizes improved rollout, telemetry, and evaluation practices to reduce the chance of similar regressions.
What happened
Anthropic investigated widespread user reports of poorer results in Claude Code and traced the behavior to three separate engineering changes that affected different traffic slices. The core API and inference layer were unaffected. All fixes are deployed as of April 20 in v2.1.116, and subscriber usage limits were reset on April 23. Anthropic wrote, "We take reports about degradation very seriously."
Technical details
Three independent changes combined to create an apparent broad regression:
- •On March 4, the default reasoning effort for Claude Code was changed from high to medium to reduce extreme latency in high mode; this decreased output quality for Sonnet 4.6 and Opus 4.6. The change was reverted on April 7.
- •On March 26, a session-cleanup change intended to clear older internal "thinking" after an hour of idle time contained a bug that made the cleanup run every turn, producing forgetful and repetitive behavior; this affected Sonnet 4.6 and Opus 4.6 and was fixed on April 10.
- •On April 16, Anthropic added a system prompt instruction to reduce verbosity; combined with other prompt changes, this harmed coding quality across Sonnet 4.6, Opus 4.6, and Opus 4.7 and was reverted on April 20.
Context and significance
This postmortem highlights the engineering tradeoffs between latency, verbosity, and reasoning effort in production model deployments. Changing defaults, session management, or system prompts can subtly shift distributional behavior and user-facing quality, especially across multiple model families and UI surfaces like Claude Code, the Claude Agent SDK, and Claude Cowork. The incident also shows how staggered rollouts and mixed-traffic effects can make correlated, but independent, changes look like a single systemic regression.
What to watch
Anthropic plans to harden rollout processes, improve telemetry and eval coverage, and make it easier for users to opt for higher reasoning effort. Practitioners using Claude Code or embedded SDKs should verify defaults in their integrations and monitor behavior after SDK or prompt updates.
Scoring Rationale
This is a notable operational postmortem that affects developers relying on Claude Code and related SDKs. It does not introduce new models or industry-wide shifts, but it highlights important production risks and engineering lessons for practitioners.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
