An update on recent Claude Code quality reports
02:00 · April 23, 2026 · Antropic Engineering Blog
We traced recent reports of Claude Code quality issues to three separate changes. Here's what happened and what we're changing.
Summary
Anthropic has published a postmortem examining three distinct changes that together produced noticeable quality degradation in Claude Code, the Claude Agent SDK, and Claude Cowork over several weeks, while leaving the underlying API unaffected. The issues surfaced at different times and affected different user cohorts, which made the pattern appear as broad but inconsistent performance loss that was initially difficult to separate from normal variation in feedback.
The first change occurred after the February release of Opus 4.6, when the default reasoning effort was lowered from high to medium to reduce occasional long-tail latency and token consumption. Internal evaluations had shown that medium effort delivered acceptable intelligence at lower cost for most tasks, yet many users experienced the model as less capable and did not adjust the setting themselves. The default was restored to high for Opus 4.7 and xhigh for Opus 4.6 on 7 April.
A second problem stemmed from an efficiency modification to prompt caching introduced on 26 March. The intended logic cleared stale thinking blocks only after an hour of inactivity, but a bug caused the clear_thinking flag to remain active for every subsequent turn. As a result, reasoning history was progressively discarded, producing the forgetfulness, repetition, and erratic tool use reported by users. The defect also triggered repeated cache misses that accelerated usage-limit consumption. It was corrected in version 2.1.101 on 10 April.
The third issue arose from a system-prompt adjustment shipped with Opus 4.7 on 16 April to curb verbosity. Although the change passed the evaluations run at the time, later ablations revealed a roughly 3 percent intelligence regression on both Opus 4.6 and 4.7. The prompt was reverted as part of the 20 April release (v2.1.116).
In response, Anthropic is expanding internal dogfooding to the exact public build, strengthening code-review tooling with additional repository context, enforcing broader per-model evaluations and soak periods for any prompt change that could affect capability, and improving auditability of system-prompt modifications. Usage limits have been reset for all subscribers.
Why it matters
This article is highly relevant for product teams and builders using Claude models, as it provides deep technical insights into prompt caching, system prompt tuning, and latency-intelligence tradeoffs. Dutch AI practitioners can apply these learnings to optimize their own LLM implementations and better understand recent anomalies in Claude's performance.
