Anthropic Claude Code Quality Reports Update: Postmortem on Regressions and Fixes
Over the past month, Anthropic investigated reports of degraded performance in Claude’s responses, tracing the issues to three distinct changes affecting Claude Code, the Claude Agent SDK, and Claude Cowork. All three issues have been resolved as of April 20 (v2.1.116), and the core API and inference layer were confirmed to be unaffected.
This post details the root causes of these regressions, the fixes implemented, and the new procedural safeguards Anthropic is adopting to prevent similar quality issues in the future.
Overview of Quality Degradation Issues
Aggregate reports of inconsistent degradation were caused by three separate, unrelated changes deployed at different times. The initial investigation was challenging because neither internal usage nor evaluations initially reproduced the identified issues, and the changes affected different slices of traffic. The issues were ultimately resolved through user feedback, which allowed Anthropic to identify and confirm the root causes.
Regression 1: Change to Claude Code's Default Reasoning Effort
Anthropic reverted a change made on March 4 that altered Claude Code's default reasoning effort from high to medium. The initial change was intended to reduce disproportionate latency and prevent the UI from appearing frozen in high-effort mode, but this tradeoff resulted in users reporting that Claude Code felt less intelligent.
When Opus 4.6 was released in February, the default reasoning effort was set to high. Following user feedback regarding occasional long-tail latencies, the default was lowered to medium. Internal testing showed that medium effort achieved slightly lower intelligence but significantly less latency for most tasks. However, after rollout, most users retained the medium default, leading to reports of reduced intelligence. Consequently, Anthropic reversed this decision on April 7, setting the default effort to xhigh for Opus 4.7 and high for all other models.
Regression 2: Caching Optimization Bug Dropping Prior Reasoning
An efficiency improvement deployed on March 26, intended to reduce the cost of resuming idle sessions, introduced a critical bug that caused Claude to appear forgetful and repetitive. The design intended to clear old thinking sections only once after a session was idle for over an hour. However, a bug caused this clearing action to execute on every turn for the remainder of the session.
This continuous clearing of thinking history meant that subsequent requests only retained the most recent block of reasoning, compounding the issue. This memory loss surfaced as forgetfulness, repetition, and odd tool choices. Furthermore, the continuous dropping of thinking blocks resulted in cache misses, which contributed to separate reports of usage limits draining faster than expected. The bug was fixed on April 10 in v2.1.101.
Regression 3: System Prompt Change to Reduce Verbosity
Anthropic reverted a system prompt instruction added on April 16, which was intended to reduce the verbosity of Claude Opus 4.7, which tends to be verbose by nature. The specific instruction added to the system prompt was:
“Length limits: keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail.”
While this change was shipped alongside Opus 4.7, subsequent ablations (removing lines from the system prompt to test impact) showed a 3% drop in intelligence for both Opus 4.6 and 4.7. The prompt was immediately reverted as part of the April 20 release.
Future Quality Assurance and Preventative Measures
To ensure similar issues are much less likely to happen again, Anthropic is implementing several structural and procedural changes to its development lifecycle:
- Increased Internal Fidelity: A larger share of internal staff will now use the exact public build of Claude Code, rather than internal testing versions.
- Enhanced Code Review: Anthropic is improving its internal Code Review tool and shipping this improved version to customers. Back-testing showed that Opus 4.7 could find the caching bug when provided complete code repositories, indicating a need for better context handling. Support for additional repositories as context for code reviews is being added.
- Tighter Prompt Controls: New tooling has been built to make system prompt changes easier to review and audit. Anthropic will run a broad suite of per-model evaluations for every system prompt change, continuing ablations to understand the impact of each line.
- Gradual Rollouts: For any change that could potentially trade off against intelligence, Anthropic will implement soak periods, a broader evaluation suite, and gradual rollouts to catch issues earlier.