Over the past month users flagged declining output from Claude Code, prompting Anthropic to trace the problem to three separate releases. The issues touched Claude Code, the Claude Agent SDK, and Claude Cowork, while the public API stayed untouched. All three bugs were patched in version v2.1.116 on April 20.

Anthropic first lowered Claude Code’s default reasoning effort on March 4 to medium, aiming to eliminate UI freezes, but users reported weaker results; high effort was reinstated on April 7 for Sonnet 4.6 and Opus 4.6. A March 26 cache change mistakenly erased reasoning history each turn, fixed April 10. An April 16 prompt tweak to curb verbosity degraded Opus 4.7 code output and was reverted April 20.

Anthropic emphasized that the public API and inference layer never suffered degradation, and as of April 23 all subscriber limits were reset. The episode illustrates how tweaking effort levels, cache policies, or system prompts can ripple through coding assistants, affecting latency, token usage, and output fidelity. With the fixes deployed, Claude Code now operates at its intended performance baseline.

To prevent similar regressions, Anthropic is tightening release gating and adding automated tests that simulate idle‑session cache behavior and effort‑level tradeoffs. Engineers will also monitor real‑time user feedback dashboards for sudden drops in code quality. These safeguards aim to keep Claude Code’s reasoning depth and responsiveness aligned with developer expectations.