Anthropic Claude Code Quality Postmortem

Anthropic has identified and resolved three distinct technical issues that caused perceived quality degradation in Claude Code, the Claude Agent SDK, and Claude Cowork. All issues were resolved as of April 20 (v2.1.116), and the company has reset usage limits for all subscribers as of April 23.

Root Causes of Quality Degradation

Anthropic traced the reported decline in performance to three separate changes implemented between March and April 2026. The API and inference layer remained unaffected throughout these events.

1. Default Reasoning Effort Reduction

On March 4, Anthropic changed the default reasoning effort for Claude Code from high to medium. This was intended to reduce long tail latencies that caused the UI to appear frozen and to maximize user usage limits.

While internal evaluations suggested medium effort provided sufficient intelligence with significantly lower latency for most tasks, user feedback indicated a noticeable drop in intelligence. Anthropic reverted this change on April 7. Current defaults are now set to xhigh effort for Opus 4.7 and high effort for all other models.

2. Session Memory and Caching Bug

On March 26, a caching optimization was introduced to reduce the cost of resuming sessions idle for over an hour by clearing old thinking sections using the clear_thinking_20251015 API header with keep:1.

A bug in the implementation caused the system to clear thinking history on every turn for the remainder of the session, rather than just once upon resumption. This led to:

  • Forgetfulness and Repetition: Claude lost memory of why it had made previous edits or tool calls.
  • Increased Costs: Continuous dropping of thinking blocks caused frequent cache misses, draining user usage limits faster than expected.

This issue was fixed on April 10 (v2.1.101). Anthropic noted that Opus 4.7 was able to identify this bug during back-testing when provided with full repository context, whereas Opus 4.6 could not.

3. Verbosity Constraints in System Prompts

On April 16, a system prompt instruction was added to reduce the verbosity of Opus 4.7, which tended to produce excessive output tokens. The specific instruction required keeping text between tool calls to $\le 25$ words and final responses to $\le 100$ words unless more detail was required.

Subsequent ablations revealed that this constraint caused a 3% drop in performance for both Opus 4.6 and 4.7. The prompt was reverted on April 20.

Future Preventative Measures

To prevent similar regressions, Anthropic is implementing the following changes to its development and deployment pipeline:

  • Internal Testing: A larger share of internal staff will now use the exact public builds of Claude Code to ensure testing environments match user experiences.
  • Code Review Tooling: Improvements are being made to the internal Code Review tool, with plans to ship the improved version to customers.
  • Prompt Governance: Anthropic is introducing tighter controls on system prompt changes, including broad per-model evaluations, continued ablations to understand the impact of individual lines, and new auditing tools.
  • Deployment Strategy: Changes involving trade-offs against intelligence will now require soak periods, broader evaluation suites, and gradual rollouts.
  • Model-Specific Gating: Guidance has been added to CLAUDE.md to ensure model-specific changes are only applied to the targeted models.

Sources

Related

  • Dispatch
    Anthropic Postmortem: Infrastructure Bugs Affecting Claude Response Quality

    Anthropic has resolved three separate infrastructure bugs that intermittently degraded Claude's response quality between August and September 2025, clarifying that model quality is never intentionally reduced due to server load.

  • Dispatch
    Claude Opus 4.8 Release Notes

    Anthropic has released Claude Opus 4.8, an upgraded model featuring improved reasoning, higher honesty, and new agentic capabilities alongside the introduction of dynamic workflows and effort controls.

  • Dispatch
    Claude Service Outage: Elevated Error Rates Across Multiple Models

    On June 23, 2026, Anthropic experienced a service disruption causing elevated error rates across claude.ai, the Claude Console, Claude API, Claude Code, and Claude Cowork.

  • Dispatch
    Claude 4 release notes / what's new

    Anthropic has released Claude Opus 4 and Claude Sonnet 4, introducing state-of-the-art coding capabilities, extended thinking with tool use, and the general availability of Claude Code.

  • Dispatch
    Designing AI-Resistant Technical Evaluations

    Anthropic details how it evolved its performance engineering take-home test through three iterations to maintain evaluation signal as Claude models increasingly outperformed human candidates.