GPT-6.1 Sol Disruption Hit Work and Codex for 8 Hours
At 08:10 UTC on October 13, OpenAI opened an investigation into degraded GPT-6.1 Sol performance. The company marked the incident resolved at 16:36 UTC. Its OpenAI incident record gives the duration as 8 hours and 26 minutes.
The affected components were ChatGPT Work and Codex Web. Codex API was also affected. So were Codex CLI and the VS Code extension. OpenAI classified the event as degraded performance, not a complete outage. Some users may have seen slow responses or intermittent errors; others may have noticed no problem.
Recovery is confirmed, but the cause is not
OpenAI identified degraded performance eight minutes after the investigation began. It applied a mitigation at 09:30 UTC, then monitored recovery before resolving the incident. The published window does not show when user impact first began. The 08:10 start time therefore does not establish that every affected request began failing then. UptimeRobot's public incident tracker records updates at 08:18 and 08:22 UTC, before the mitigation.
The incident record names no root cause or affected-request percentage. It gives no regional pattern. OpenAI published no error codes or rollback details. The company also did not say whether queued or interrupted tasks resumed automatically. Its disclosure confirms degradation on the listed surfaces, but does not establish lost work or duplicated actions. Nor does it establish a model-quality regression or security breach.
A separate user report in the OpenAI Codex GitHub tracker described Windows Codex errors beginning on October 12. The report included HTTP 503 responses and "Selected model is at capacity" messages. By October 13, its author had counted 59 capacity messages and 25 HTTP 503 responses. OpenAI has not linked the report to the GPT-6.1 Sol incident.
Separately, users had reported scheduled-task errors in ChatGPT Work Mode. OpenAI has not linked that issue to the GPT-6.1 Sol incident. On October 13, the company also logged a separate issue involving the creation of new branches in Codex Dot. It marked that incident resolved after the affected services recovered.
Check task state before trying again
For a failed chat response, OpenAI's troubleshooting guidance recommends checking the status page, then refreshing or restarting the app. For transient failures, it recommends trying again. Agent work calls for one extra check: inspect the original task and its outputs before resending the same instruction.
A timed-out interface does not prove an operation failed. Check whether a file changed. Confirm whether a command completed or a pull request was created. For API mutations, reconcile the application record with any idempotency key before retrying. Rerun only work confirmed missing.
Teams can compare local logs with the provider-reported 08:10-16:36 UTC window, while treating the actual impact start as unknown. For unresolved problems, retain the model and surface. Keep the task or conversation ID with the report. Add the exact UTC timestamp and last confirmed action for Support. OpenAI's incident record lists the event as resolved.
