Anthropic posted a postmortem on recent Claude Code quality complaints, stating that the apparent degradation traced back to three separate changes rolled out between early March and mid-April. The company affirms that the API was not impacted and that all three issues were resolved by April 20, with usage limits for all subscribers reset as of April 23.
The postmortem describes three distinct problems:
1. On March 4, Claude Code’s default reasoning effort was changed from `high` to `medium` to reduce long latencies that could make the UI appear frozen. Anthropic now calls that “the wrong tradeoff” and reverted it on April 7. The change affected Sonnet 4.6 and Opus 4.6. 2. On March 26, Anthropic shipped a change intended to clear older thinking from sessions idle for more than an hour. A bug caused that clearing to repeat on every turn for the rest of the session, which the company links to reports that Claude seemed forgetful and repetitive. That issue affected Sonnet 4.6 and Opus 4.6 and was fixed on April 10. 3. On April 16, the company added a system-prompt instruction to reduce verbosity. In combination with other prompt changes, Anthropic states that it hurt coding quality and was reverted on April 20. That affected Sonnet 4.6, Opus 4.6, and Opus 4.7.
Anthropic’s account suggests that the combined effect looked like broad, inconsistent degradation because the changes hit different parts of traffic on different schedules. The company also states that early investigations were complicated by ordinary variation in feedback and by the fact that internal usage and evals did not initially reproduce the issue.
Default reasoning effort changed, then reverted
Anthropic says it originally set Opus 4.6 in Claude Code to `high` reasoning effort when it launched that model in February. After users complained that the model would sometimes think for so long that the interface appeared frozen, the company switched the default to `medium`.
The post argues that medium effort delivered “slightly lower intelligence” but much lower latency for most tasks, while also reducing usage-limit hits. Even so, many users reportedly still preferred higher intelligence over speed. After more feedback, Anthropic reversed course on April 7. The company now defaults to `xhigh` effort for Opus 4.7 and `high` effort for all other models.
A caching change appears to have dropped reasoning history
A second issue involved prompt caching and what Anthropic describes as an optimization for sessions idle for more than an hour. The company states that the intended behavior was to clear older thinking once, reduce the number of uncached tokens sent back to the API, and then resume sending full reasoning history.
Instead, the implementation allegedly kept clearing thinking on every turn after a session crossed the idle threshold. Anthropic links that bug to reports of forgetfulness, repetition, and odd tool choices, and also to usage limits draining faster than expected because the repeated clearing caused cache misses.
The company notes that the bug sat at the intersection of Claude Code’s context management, the Anthropic API, and extended thinking. It also claims the issue passed multiple code reviews, tests, verification steps, and dogfooding before being discovered. Anthropic says it later back-tested Code Review against the offending pull requests using Opus 4.7 and found that, when given enough repository context, Opus 4.7 identified the bug while Opus 4.6 did not.
That fix landed on April 10 in v2.1.101.
A verbosity tweak had an outsized effect
A third change involved a system-prompt instruction meant to reduce verbosity for Claude Opus 4.7, which Anthropic had already described elsewhere as relatively chatty. The line added to the prompt read:
> “Length limits: keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail.”
Anthropic states that internal testing did not show regressions at first, but broader ablations later found a 3% drop for both Opus 4.6 and 4.7. The company reverted the prompt on April 20.
What Anthropic plans to change
To avoid similar regressions, Anthropic says it will take several steps: more internal staff will use the exact public build of Claude Code, the company will improve its internal Code Review tool and ship that to customers, and every system-prompt change will get a broader per-model eval suite.
The company also mentions new tooling for prompt review and audit, updated guidance in CLAUDE.md so model-specific changes stay gated to the intended model, and soak periods plus gradual rollouts for changes that might trade off against intelligence.
Anthropic recently created @ClaudeDevs on X to explain product decisions in more depth, and it plans to carry the same updates into centralized GitHub threads. The company closes by thanking users who filed feedback or shared reproducible examples online, which it credits with helping identify and fix the problems.
Source: Anthropic Engineering
