# anthropic-claude-code-quality-postmortem-2026-04-23

## Veille

Claude Code Quality Post-Mortem March-April 2026 — Three Caching/Reasoning/Prompt Incidents — Anthropic Engineering Blog

## Titre Article

An Update on Recent Claude Code Quality Reports

## Date

2026-04-23

## URL

https://www.anthropic.com/engineering/april-23-postmortem

## Keywords

post-mortem, Claude Code, quality degradation, reasoning effort, caching bug, prompt caching, system prompt, extended thinking, intelligence regression, Claude Agent SDK, Claude Cowork, Opus 4.7, rollback, quality process, ablation evaluation

## Authors

Anthropic

## Ton

**Profile**: Engineering post-mortem post published by Anthropic, technical-transparent register, intermediate-advanced level.

**Description**: Anthropic adopts a tone of deliberate technical transparency, typical of Silicon Valley engineering post-mortems. The style is factual, chronological, and detailed, with a clear intent to be accountable to users. The article methodically distinguishes three separate incidents, each documented with its dates, root cause, impact, and resolution. The vocabulary is technical yet accessible (reasoning effort, prompt caching, ablation testing). Authority stems from the position of the product's own publisher — Anthropic acknowledges its mistakes and details corrective measures. The target audience is the developer community using Claude Code, the Claude Agent SDK, and Claude Cowork, as well as the broader tech community concerned with product reliability.

## Pense-betes

- Three distinct incidents degraded Claude Code's quality between March and April 2026 — the API itself was never affected, indicating the issues were specific to Claude Code's scaffolding/harness
- Incident 1 (March 4 - April 7): the change of the default reasoning level from `high` to `medium` to reduce latency caused a perceived drop in intelligence. Despite UX iterations (notices, selector), users stayed on `medium`. Resolved by defaulting to `xhigh` for Opus 4.7 and `high` for the other models
- Incident 2 (March 26 - April 10): a critical prompt caching bug — the `clear_thinking_20251015` header with `keep:1`, meant to run only once on sessions inactive for >1h, triggered on every subsequent turn, progressively erasing the reasoning context. Consequences: cascading cache misses, Claude "forgetful and repetitive," accelerated usage quota consumption
- The caching bug was difficult to detect because unrelated internal experiments and display changes masked the issue during testing — a classic testability debt pattern
- Notable fact: Opus 4.7 with the Code Review tool identified the bug retrospectively when given the full repository context, whereas Opus 4.6 could not — a concrete illustration of the capability jump between model versions
- Incident 3 (April 16-20): a verbosity constraint in the system prompt ("keep text between tool calls <= 25 words, final responses <= 100 words") caused a 3% drop in intelligence measured by ablation testing, for both Opus 4.6 and 4.7 models
- The three incidents reveal a recurring tension in coding agent engineering: optimizing latency/cost vs. preserving reasoning quality — every shortcut carries a cognitive cost for the model
- User reports via `/feedback` and reproducible examples were credited as critical to identifying and resolving these issues — validation of the community feedback model
- Announced process improvements: increased internal use of public builds, per-model evaluations, ablation documentation, stabilization periods, phased rollouts, creation of the @ClaudeDevs account on X
- All issues resolved by April 20 (v2.1.116), with a reset of usage limits for all subscribers on April 23 — an unusual commercial gesture reflecting the perceived severity of the impact

## RésuméDe400mots

In this engineering post-mortem, Anthropic documents three distinct incidents that degraded the perceived quality of Claude Code, the Claude Agent SDK, and Claude Cowork between March and April 2026, while specifying that the underlying API was never affected.

The first incident (March 4 - April 7) involved a configuration change to the default reasoning level, switched from "high" to "medium" to resolve interface freezing issues caused by extended thinking in high mode. Internal tests showed that medium mode offered "slightly lower intelligence with significantly reduced latency." However, users quickly reported that Claude seemed "less intelligent." Despite several design iterations (notifications, effort selector), users retained the medium default. Anthropic ultimately reversed its decision by switching to "xhigh" level for Opus 4.7 and "high" for the other models.

The second incident (March 26 - April 10) is the most technical and the most damaging. A prompt caching optimization intended to clean up old thinking sections from sessions inactive for more than an hour contained an implementation flaw. The API header `clear_thinking_20251015` with the `keep:1` parameter was meant to run only once but triggered on every subsequent turn, progressively erasing Claude's reasoning context. This caused cascading cache misses, making Claude "forgetful and repetitive" and depleting usage quotas faster. The bug proved difficult to detect because unrelated internal experiments masked the issue. Notably, it was Opus 4.7's Code Review tool, fed with the full repository context, that identified the bug retrospectively — Opus 4.6 had not been able to.

The third incident (April 16-20) resulted from an instruction added to the system prompt limiting verbosity (text between tool calls capped at 25 words, final responses at 100 words). Internal tests had detected no regression, but broader ablation tests revealed a 3% intelligence drop for both Opus 4.6 and Opus 4.7.

All issues were resolved by April 20 with version 2.1.116. Anthropic reset usage limits for all subscribers on April 23. The company announced several process improvements: increased internal use of public builds, per-model evaluations, systematic ablation testing, stabilization periods, phased rollouts, and the creation of the @ClaudeDevs account on X for more detailed product communication.

## GrapheDeConnaissance

- Anthropic —publie→ post-mortem qualité Claude Code avril 2026 (DOCUMENT, 0.98)
- Trois incidents de dégradation qualité —observé_dans→ Claude Code (TECHNOLOGIE, 0.98)
- Niveau de raisonnement medium par défaut —remplace→ niveau de raisonnement high par défaut (CONCEPT, 0.97)
- Changement reasoning effort —réduit→ intelligence perçue par les utilisateurs (CONCEPT, 0.95)
- Bug prompt caching —permet→ cache misses en cascade et perte de contexte (CONCEPT, 0.97)
- Anthropic —affirme_que→ l'en-tête clear_thinking_20251015 devait s'exécuter une seule fois sur les sessions inactives (AFFIRMATION, 0.95)
- Anthropic —affirme_que→ l'en-tête clear_thinking_20251015 se déclenchait à chaque tour suivant (bug) (AFFIRMATION, 0.97)
- Opus 4.7 —résout→ identification rétrospective du bug de cache (EVENEMENT, 0.93)
- Opus 4.7 —surpasse→ Opus 4.6 (TECHNOLOGIE, 0.9)
- Contrainte verbosité —mesure→ baisse de 3% d'intelligence (tests d'ablation) (MESURE, 0.95)
- Anthropic —publie→ réinitialisation des limites d'utilisation de tous les abonnés (23 avril) (EVENEMENT, 0.95)
- Anthropic —a_créé→ @ClaudeDevs (ORGANISATION, 0.9)
- Remontées utilisateurs /feedback —permet→ identification et résolution des bugs (CONCEPT, 0.92)
- Anthropic —affirme_que→ l'API Claude n'a jamais été affectée par les trois incidents (AFFIRMATION, 0.98)

---
Canonical: https://www.thekb.eu/en/fiches/anthropic-claude-code-quality-postmortem-2026-04-23/