Anthropic has published a postmortem identifying three separate product-layer changes that degraded performance across Claude Code, the Claude Agent SDK and Claude Cowork between early March and late April, closing out weeks of user complaints about lower-quality outputs and faster-than-expected usage limit depletion.
The company's Claude API was not affected, and Anthropic stressed that the underlying model weights had not been altered.
The first issue dates to 4 March, when Anthropic reduced Claude Code's default reasoning effort from "high" to "medium" to address latency problems that made the interface appear frozen during extended thinking.
Users immediately reported that the tool felt less intelligent, and after sustained pushback, the company reversed the change on 7 April.
"This was the wrong tradeoff," Anthropic said.
The latest Claude Code build now defaults to "xhigh" reasoning effort on Opus 4.7 and "high" on all other models.
The second issue involved a cache optimisation shipped on 26 March that was designed to clear old reasoning blocks from sessions idle for more than an hour.
A bug caused the clearing to repeat on every subsequent turn rather than firing once, stripping the model of its reasoning history and making it appear forgetful and repetitive.
The continuous deletion also caused frequent cache misses, which Anthropic said was likely the root cause of reports that usage limits were draining too quickly.
The fix for Sonnet 4.6 and Opus 4.6 shipped on 10 April.
The bug had slipped past multiple human and automated code reviews, unit tests, end-to-end tests and internal dogfooding, in part because a separate internal experiment that changed how thinking was displayed had suppressed the symptoms in most sessions used by Anthropic's own engineers.
The third issue arose on 16 April, when a system prompt change instructed models to keep text between tool calls to 25 words or fewer and final responses to 100 words unless the task required more detail.
After shipping the change with Opus 4.7, expanded ablation testing revealed a 3% drop in coding evaluation scores across both Opus 4.6 and Opus 4.7, and the adjustment was reverted on 20 April.
Because each change affected a different slice of users on a different schedule, the combined effect appeared as broad, inconsistent degradation that was initially difficult to distinguish from normal variation in feedback.
An external audit of more than 6,800 Claude Code session files by Stella Laurenzo, a senior director in AMD's AI group, added to the pressure on Anthropic to investigate further.
Anthropic said it will require more internal staff to use public Claude Code builds, apply stricter evaluation to system prompt changes, improve its Code Review tool and create a dedicated @ClaudeDevs account on X to explain product decisions.
The company has reset account usage levels for all customers as of 23 April.
"This isn't the experience users should expect from Claude Code," Anthropic said.
The recap
- Three overlapping changes degraded Claude products' output quality.
- System prompt change produced a 3% performance drop.
- Anthropic will increase testing and add a @ClaudeDevs X account.
