US - Agents - Some Customers May Experience Errors When Using Claude Sonnet 4.6

Update timeline

1. investigating Aug 19, 2026, 03:08 PM UTC

   We are investigating an issue that may affect some customers using Claude Sonnet 4.6 in Agents in the US region. Our engineering team is actively working to understand the issue and will share further updates as more information becomes available.

2. investigating Aug 19, 2026, 04:34 PM UTC

   We are investigating an issue that may affect some customers using Claude Sonnet 4.6 in Agents in the US region. Our engineering team is actively working to understand the issue and will share further updates as more information becomes available.

3. monitoring Aug 19, 2026, 04:40 PM UTC

   We have mitigated the issue. Our engineering team is actively morning and will share further updates as more information becomes available.

4. monitoring Aug 19, 2026, 04:42 PM UTC

   We have mitigated the issue. Our engineering team is actively monitoring and will share further updates as more information becomes available.

5. resolved Aug 19, 2026, 04:55 PM UTC

   The issue has been resolved.

6. postmortem Aug 31, 2026, 12:41 PM UTC

   ## Customer Impact Between August 19, 2026 at 1:36 pm UTC and August 19, 2026 at 3:53 pm UTC, a subset of customers received errors when using Claude Sonnet 4.6 in Agents. The impact lasted approximately 2 hours and 17 minutes. The impact was to customers using Agents in the US region. Errors were also observed for Claude Opus 4.6 and Claude Opus 4.5, which are used at lower volume. --- ## Root Cause As part of a planned infrastructure migration, we moved the platform service that routes model requests for Agents onto a new configuration delivery system, region by region. The new configuration source was missing the routing entries for three Claude models — Claude Sonnet 4.6, Claude Opus 4.6, and Claude Opus 4.5. Without those entries, the service could not resolve a valid destination for requests to those models, and rejected them with errors. Other models were unaffected and continued to serve normally throughout. --- ## Detection The issue was detected through customer escalations on August 19, 2026 at 2:55 pm UTC — approximately 1 hour and 19 minutes after the first affected request. Our engineering team scoped the issue to specific Claude models and began investigation. Public status communication began at 3:08 pm UTC. Our alert monitoring is based on aggregate error rates. Although nearly all requests to the three affected models were failing, those models represented a small share of overall traffic in the region, so the aggregate signal did not cross our alerting thresholds and the issue was not raised automatically. This is the detection gap addressed in the follow-up below. --- ## Response At 3:25 pm UTC, the incomplete configuration source was identified as the cause, and a fix was started. At 3:47 pm UTC, deployment of the fix was in progress, and the service was actively monitored as the change rolled out. By 3:59 pm UTC, service logs confirmed the issue was mitigated, and at 4:00 pm UTC the failure rate was confirmed at 0%. The incident was marked mitigated at 4:40 pm UTC, and full resolution was declared at 4:55 pm UTC after continued monitoring and customer confirmation that the service was working as expected. --- ## Follow-Up The infrastructure migration has been completed across all regions and routing configuration now draws from a single source, removing the mismatch that caused this incident so it cannot recur. Automated checks are being introduced to continuously verify every supported model in every region, so an unavailable model is detected and alerted immediately — including in lower-traffic regions.