OpenAI's Costs API exported incorrect cost data for four days before anyone flagged it. The fault sat with OpenAI's billing pipeline, not with your integration, and the fix took 2h 20m to land.
Started
04:42 UTC
Duration
Lasted 2h 20m
Source
IsDown
Next time
OpenAI is degraded. Is your product working?
An incident on their side does not always mean an outage on yours, and the only way to know is to already be watching your own endpoints. Set that up in under a minute and you will be notified the next time your own site is affected. And there will be a next time.
50 monitors for your own websites. No card.
What happened?
OpenAI logged one incident on 10 November 2026, starting at 04:42 UTC and lasting 2h 20m. The issue was not a request failure or an outage in the usual sense. OpenAI's own update says it identified "incorrect cost data in Costs API exports published between November 6 and 9," meaning anyone pulling cost figures from that API during that window got wrong numbers without any error being thrown. The single update posted is a "monitoring" status, OpenAI's own term for saying it is watching a fix take effect rather than declaring the problem solved. No components are listed as affected on OpenAI's status page for this entry, which is consistent with a backend data correction rather than a service disruption that would touch multiple visible systems. OpenAI said it was "working to correct the affected records" and would confirm once recovery was verified, but no further update appears in the record after the initial post. That means this page cannot tell you whether the correction actually completed or what records were restored, because OpenAI did not file a closing update. No user reports were filed during the window, which fits an issue like this: if your bills looked a little off for a few days, you would have no way to tell from normal usage that the numbers themselves were wrong rather than your spend simply being higher. OpenAI's history shows 101 incidents in the past 90 days with a median duration of 1 hour 46 minutes, so this one ran a little longer than typical. Across OpenAI's full history, incidents average 15.3 per month, so a single-incident day like this one is on the lighter end.
Learning
This incident did not break any endpoint you were calling. It corrupted the figures a downstream reporting tool would show you, which is a different kind of failure to the request timeouts or 500s that uptime checks are built to catch. UptimeRobot's own 30 day figures show 30 incidents logged against OpenAI with a median duration of 1h 58m, and 29 of those resolved, numbers that come from watching OpenAI's status page and the provider's own component history, not from pinging an endpoint and getting a response back. A clean response from the Costs API would have told you nothing was wrong, because the API was up and returning data, just the wrong data. That is the gap between "the service answered" and "the service answered correctly," and it is a gap that external endpoint checks cannot close on their own. If you rely on OpenAI's cost exports for billing, budgeting or alerts, treat any figures from 6 to 9 November with suspicion until OpenAI confirms the correction, since no confirmation appears on the record. Your monitoring covers your half, UptimeRobot now watches the provider's half too, so next time you know which side broke without guessing.