2026-02-06: Inference error rate for anthropic model in europe-west3 region exceeds SLO

Inference error rate for anthropic model in europe-west3 region exceeds SLO (Severity 4)

Problem: Non-streaming inference requests to the anthropic model in the ai-gateway for the europe-west3 region are returning errors at a rate of 11.62%.

Impact: 11.62% of non-streaming inference requests to the anthropic model in the ai-gateway for the europe-west3 region are returning errors, which violates the service level objective and affects the Duo Agentic Platform (DAP).

Causes: Prompts sent to the Anthropic proxy endpoints in the europe-west3 region are exceeding the allowed length. The proxy currently returns a 502 Bad Gateway error instead of the original 400 Bad Request error from Anthropic, which misclassifies validation errors as service failures. The affected model is claude-sonnet-4-5-20250929.

Response strategy: We have proposed updating the proxy to surface the original 400 Bad Request error from Anthropic instead of a 5XX error. A merge request has been opened to implement this fix.

This ticket was created to track INC-7177, by incident.io 🔥