incidents.si
INC-2025-0015SEV3 · Mediumoutageconfirmed

Three infrastructure bugs degrade Claude’s responses for weeks

Anthropic reported that three overlapping infrastructure bugs lowered the quality of Claude’s answers between August and early September 2025, and explained why they were slow to detect.

Occurred
Disclosed
Published
Organization
Anthropic
System
Claude API and Claude Code
Impact
Intermittently degraded responses from early August to mid-September; 16% of Sonnet 4 requests affected at the worst hour
Root cause
A routing error, a serving misconfiguration and a compiler bug, overlapping in time

Timeline

  • 2025-08-05. A routing error starts sending some Sonnet 4 requests to servers set up for a different context window. It affects 0.8% of requests at first.
  • 2025-08-25. Two further bugs start: one corrupts output, one affects token selection.
  • 2025-08-29. A load balancing change increases the misrouting.
  • 2025-08-31. In the worst hour, 16% of Sonnet 4 requests are misrouted.
  • 2025-09-02 to 2025-09-12. The second and third bugs are rolled back.
  • 2025-09-04. A fix for the routing error is deployed. The rollout completes by 2025-09-18.

Trigger or entry point

Routine infrastructure changes: a routing change, a serving configuration change and a change to how tokens are selected.

Failure mechanism

The first bug sent requests to the wrong type of server. The second, a misconfiguration, sometimes produced corrupted output. The third exposed a compiler bug in the code that picks the next token. Each had different symptoms on different platforms, which made the overall picture confusing.

Impact

About 30% of Claude Code users in the period had at least one message misrouted. On third-party platforms the share of affected requests was far lower. This entry is filed under outage because the service stayed up but its quality dropped.

Response

Anthropic fixed or rolled back all three changes and published a detailed postmortem.

Mitigations

Anthropic names four changes:

  • more sensitive evaluations,
  • quality evaluations that run continuously on production systems,
  • faster tooling to debug user reports without weakening privacy controls,
  • tests that detect unexpected characters in output.

Anthropic also explains why detection was slow: its evaluations did not capture the degradation, and its privacy controls limit engineers’ access to user interactions.

Sources

  1. A postmortem of three recent issuesprimary · 2025-09-17

Corrections

No corrections have been made to this entry.

Report a correction