← Back to feed

Anthropic published a public engineering postmortem on April 23 tracing months of Claude Code quality complaints to three separate changes that overlapped in time and affected different slices of traffic — making the degradation pattern look broad and inconsistent when it was actually the sum of three distinct bugs.

All three issues are resolved as of April 20 (v2.1.116). The API was not impacted at any point.

The Three Changes

1. Reasoning effort downgraded, March 4 — reverted April 7

When Opus 4.6 launched in Claude Code in February, the default reasoning effort was set to high. Shortly after, users reported occasional very long think times that froze the UI and consumed disproportionate tokens. On March 4, Anthropic silently switched the default to medium effort.

Users noticed the intelligence drop and said so. Anthropic shipped UI changes to make the effort setting visible — inline selectors, startup notices, restoring the “ultrathink” keyword — but most users stayed on medium. The company reverted the default on April 7 after “hearing feedback from more customers.”

The decision is described in the postmortem as “the wrong tradeoff.” All models now default to high or xhigh effort: Opus 4.7 defaults to xhigh, all others to high.

Affected models: Sonnet 4.6, Opus 4.6.

2. Caching bug that dropped reasoning history, March 26 — fixed April 10

Claude normally keeps its chain-of-thought in conversation history so subsequent turns can see the reasoning behind prior actions. On March 26, an efficiency change was meant to prune old thinking sections from idle sessions (over one hour inactive) — since those sessions would be a cache miss anyway.

A bug caused the pruning to run on every turn for the rest of a session, not just the first turn after an idle gap. Claude appeared to forget its earlier reasoning immediately after any thinking-heavy exchange, becoming “forgetful and repetitive” within minutes of active use. Anthropic identified and fixed the bug on April 10.

Affected models: Sonnet 4.6, Opus 4.6.

3. Verbosity reduction prompt hurt coding quality, April 16 — reverted April 20

On April 16, a system prompt instruction was added to reduce response verbosity. In isolation, this was benign. Combined with other prompt changes already in place, it degraded coding output quality. The change was reverted four days later.

Affected models: Sonnet 4.6, Opus 4.6, Opus 4.7.

Why It Looked Worse Than Each Part

Each change ran on a different schedule, affected a different model subset, and hit different traffic patterns. The combined effect looked like “broad, inconsistent degradation” across Claude Code. Internal evaluations and Anthropic’s own usage initially did not reproduce the issues — partly because the bugs required specific session states (idle gaps, active reasoning sessions, specific prompt combinations) to surface reliably.

The March 4 change was in place for 33 days before reversal. The March 26 caching bug ran for 15 days. The April 16 verbosity change lasted four days.

Response

Anthropic is resetting usage limits for all Claude Code subscribers as of April 23. The postmortem commits to changes in how prompt changes are tested against session-state-dependent behaviors and how user quality signals are weighted against internal eval results during rollouts.

The Axios report accompanying the postmortem characterized Anthropic’s Q1 problems as spanning “nearly every part of its business — product quality, pricing, security and capacity.” Forbes reported that enterprise security teams had flagged Claude Opus models for introducing vulnerabilities into production code. The postmortem addresses only the Claude Code product layer and does not speak to the security claims or capacity constraints directly.