Anthropic’s Reliability Wakeup
- Anthropic said on April 23 that three separate changes degraded Claude Code, the Claude Agent SDK, and Claude Cowork after weeks of user complaints about weaker coding responses. - The company said it had lowered Claude Code’s default reasoning effort from “high” to “medium,” shipped a cache bug that dropped prior thinking, and added a brevity prompt. - Anthropic also began offering Claude Design in research preview across paid plans, with Enterprise defaulted off, as buyers weigh reliability and admin controls more closely. (anthropic.com) (support.claude.com)
Anthropic said three separate product changes caused the recent drop in Claude Code quality after a month of complaints from developers. (anthropic.com) In its April 23 postmortem, Anthropic said Claude Code’s default reasoning effort had been reduced from “high” to “medium,” which made answers less thorough for some coding tasks. (anthropic.com) The company also said a cache-clearing bug caused Claude Code to lose parts of its prior “thinking” history, which made the tool seem forgetful and repetitive during longer sessions. (anthropic.com) A third change added a brevity instruction that shortened responses across Claude Code, the Claude Agent SDK, and Claude Cowork, and Anthropic said that hurt coding quality in practice. (anthropic.com) Anthropic said it has already reversed or fixed those changes and will add stricter review for quality-sensitive updates, including stronger checks before changing defaults that affect coding behavior. (anthropic.com) The company also said it would reset usage limits for subscribers affected by the degraded experience, after users had reported burning through quotas while getting weaker output. (anthropic.com) (aol.com) The episode landed as Anthropic rolled out Claude Design in research preview for Pro, Max, Team, and Enterprise plans. Anthropic’s help center says the feature is default off for Enterprise customers. (support.claude.com) Anthropic’s help pages describe Claude Design as a system that pulls reusable colors, typography, components, and patterns from a company’s existing assets so new work matches its brand rules. (support.claude.com) Anthropic has used conservative defaults on other workplace features too: its help center says Claude Cowork launched in research preview for paid plans with admin controls so organization owners can disable it. (support.claude.com) The immediate test for Anthropic is whether the fixes hold and whether enterprise customers accept new features only when reliability, limits, and admin controls are spelled out in advance. (anthropic.com) (support.claude.com)
