AI Research Atlas

A postmortem of three recent issues

Anthropic · 17 September 2025

Anthropic traced degraded Claude answers in August 2025 to three overlapping infrastructure bugs, one misrouting up to 16% of Sonnet 4 requests.

There were three bugs. (1) Context-window routing sent Sonnet 4 requests to 1M-context servers (0.8% of requests from 2025-08-05, 16% at the worst hour on 2025-08-31). (2) A TPU-server misconfiguration corrupted token generation from 2025-08-25. (3) An approximate top-k XLA:TPU compiler miscompilation. Fixes reached Bedrock by 2025-09-18.

Date
Wednesday, 17 September 2025
Lab
Anthropic
Kind
infra
Access
closed API

Anthropic says it serves Claude on AWS Trainium, NVIDIA GPUs and Google TPUs, so each platform needs separate validation. About 30% of Claude Code users hit at least one misrouted message in the period.

Sources

  1. www.anthropic.com/engineering/a-postmortem-of-three-recent-issues

This record was checked against its sources on 6 October 2026. How we check

Related