Anthropic Publishes Claude Code Postmortem — Three Bugs Silently Degraded Quality for Weeks
Anthropic published a detailed postmortem revealing that three separate changes between March and April 2026 caused Claude Code, the Agent SDK, and Claude Cowork to feel noticeably less capable. Users reported Claude 'forgetting' context, repeating itself, and draining usage limits faster — all caused by a reasoning effort downgrade, a caching bug, and an aggressive verbosity prompt.
Original sourceAnthropic today published an engineering postmortem for a period of degraded Claude Code performance that affected users of Sonnet 4.6, Opus 4.6, and Opus 4.7 between early March and April 20, 2026. The post is notable for its candor: three separate changes, each made independently, compounded into what users experienced as a broadly intelligent model becoming forgetful, repetitive, and strangely conservative.
**The Three Root Causes**
First, on March 4, Anthropic switched the default reasoning effort from "high" to "medium" to reduce latency. Users complained the change hurt perceived intelligence quality — a real tradeoff that should have been announced rather than silently deployed. Second, on March 26, a caching bug caused Claude to clear its thinking history on every turn rather than once per session, explaining the repetitive, context-losing behavior many users noticed. Third, on April 16, system instructions were added limiting responses to "≤25 words" between tool calls — a verbosity prompt that degraded coding quality by a measured 3%.
**The Impact**
Collectively, these changes produced reports of Claude feeling "less intelligent," making unusual tool choices, and behaving inconsistently. Some users also saw usage limits drain faster than expected — likely related to the caching bug causing redundant processing. The Hacker News post about the postmortem reached 322 points, reflecting how much developer trust had been eroded by the silent degradation.
**Remediation**
All issues were resolved by April 20 in v2.1.116. Anthropic reset usage limits for all subscribers and committed to broader internal testing, enhanced Code Review tooling, tighter system prompt controls, and gradual rollouts for any changes that affect intelligence. The postmortem sets a meaningful precedent for transparency in AI product quality incidents — something the industry has often avoided.
Panel Takes
The Builder
Developer Perspective
“Respect for publishing a real postmortem rather than a vague 'we've improved quality' tweet. The caching bug explains so much — the repetitive behavior and context loss were driving me insane. The commitment to gradual rollouts for intelligence-affecting changes is exactly what enterprise users need before trusting Claude in production pipelines.”
The Skeptic
Reality Check
“Three separate changes degraded a production AI tool for six weeks, and users had no idea why they were paying for a worse product. The reasoning effort downgrade alone was a decision that should have been a changelog entry, not a silent deploy. 'Commitment to gradual rollouts' means nothing without a published policy — where's the SLA?”
The Futurist
Big Picture
“This incident reveals a structural problem: AI products can silently degrade in ways that are hard to detect because intelligence quality has no objective metric users can monitor. The postmortem format is a step toward AI product incident culture — the same rigor we apply to infrastructure outages needs to apply to capability regressions.”