An update on recent Claude Code quality reports
Summary (EN)
Anthropic published an engineering update on April 30 explaining a cluster of quality problems that some users had reported in Claude Code and related products over the previous month. The company says the issue did not come from model degradation at the API or inference layer. Instead, it traced the complaints to three separate product-layer changes affecting Claude Code, the Claude Agent SDK, and Claude Cowork. First, on March 4, Anthropic lowered Claude Code’s default reasoning effort from high to medium to reduce latency, but later concluded that this was the wrong tradeoff and reverted it on April 7 after users signaled that they preferred higher default intelligence. Second, a March 26 change meant to clear stale thinking from sessions idle for over an hour introduced a bug that repeated the reset every turn, making the model appear forgetful and repetitive until Anthropic fixed it on April 10. Third, a system-prompt instruction added on April 16 to reduce verbosity interacted badly with other prompt changes and hurt coding quality, leading to a rollback on April 20. Anthropic says all three issues were resolved by version 2.1.116 and that subscriber usage limits were reset. The announcement matters because it provides unusually concrete evidence that product-layer choices can materially alter perceived model quality even when the underlying model remains unchanged. The publication and disclosure are both within the past 24 hours.
Summary (ZH)
Anthropic 于 4 月 30 日发布了一篇工程说明，回应过去一个月部分用户对 Claude Code 及相关产品“质量变差”的反馈。公司明确表示，问题并非来自底层 API 或推理层模型能力下降，而是源于产品层连续发生的三项改动，分别影响了 Claude Code、Claude Agent SDK 和 Claude Cowork。第一项变更发生在 3 月 4 日，当时 Anthropic 为了降低部分用户在高推理模式下遇到的延迟问题，把 Claude Code 的默认 reasoning effort 从 high 调低到 medium；但后来公司认为这是错误权衡，并在 4 月 7 日回滚，因为用户更希望默认获得更高智能水平。第二项问题来自 3 月 26 日，为减少长时间闲置会话恢复时的延迟，系统引入了清理旧思考痕迹的逻辑，但 bug 导致该清理在后续每一轮都重复触发，使模型表现得健忘且重复，直至 4 月 10 日才修复。第三项问题出现在 4 月 16 日，当时新增了一条减少冗长度的 system prompt 指令，但它与其他提示层改动叠加后损害了编码质量，最终于 4 月 20 日撤回。Anthropic 称三项问题均已在 v2.1.116 中解决，并重置了订阅用户的使用额度。此事的重要性在于，它清晰展示出 AI 产品体验会受到默认设置、会话记忆逻辑和提示词层改动的显著影响，即便底层模型本身没有变化。文章发布与问题说明本身都发生在过去 24 小时内。
Source