An update on recent Claude Code quality reports

Source: Anthropic Engineering Date: 2026-04-23 URL: https://www.anthropic.com/engineering/april-23-postmortem

Summary

Anthropic’s second public postmortem, covering three overlapping Claude Code quality regressions in March–April 2026: a reasoning-effort default change (high → medium on March 4, reverted April 7), a caching bug that cleared reasoning history on every turn rather than once per idle session (shipped March 26), and a system prompt constraint (“≤25 words between tool calls”) that caused a 3% intelligence degradation (introduced April 16, reverted April 20). Stated improvements include expanded per-model eval suites for system prompt changes and gradual rollouts with soak periods.

Implications

The reliability and postmortem thread. This is the September 2025 postmortem pattern repeating: multiple unrelated bugs creating compounded apparent degradation, with slow detection. The caching bug is notable — a prompt caching optimization introduced a stateful bug that caused progressive context erosion per session, which is exactly the failure mode that context engineering (just-in-time retrieval, context resets) is designed to survive.

System prompt constraints as a regression vector. A 25-word limit on inter-tool text causing a measurable 3% intelligence drop is a clear signal that system prompt constraints on model behavior require the same eval rigor as model changes. Any shop that ships system prompt updates without per-model ablation testing is exposed to the same regression path.

Reasoning effort as a tunable parameter. The public documentation that Claude Code has a reasoning effort setting (high/medium) and that it meaningfully affects user-perceived intelligence is new operator-level information. Shops running Claude Code at scale should audit whether their deployment defaults match their intelligence requirements.