← Back to OpenAI's check readings
We started checking OpenAI on August 20, 2026, which is 79 days ago. Everything below covers that period, not the full 90 days a longer-tracked service would show.
Email me when OpenAI reports an incident on its status page, and again when it reports it resolved.
The last 24 hours
What happened most recently, on one time axis: our own probe readings by surface, how many people reported a problem, and what OpenAI's own status page declared. Below it, any public chatter we saw about OpenAI and the extra probe it triggered.
The last 24 hours on one axis: what our probes read for each surface we check, how many people reported a problem, and what OpenAI's own status page declared. Three records side by side; none of them changes another.
- Elevated error rates across paid ChatGPT plans: minor incident, from 22:30 UTC to 00:45 UTC.
Each check cell is the worst reading across every region in that quarter hour: one region that did not answer colours the cell, never an average. An amber cell is a reading that answered, but at three times or more its 7-day median for three checks running: slow is not down, and we say which. A filtered reading stays neutral, because a refused check is not a failed check. Report bars count reports as submitted; we do not check who sent them, and they never change our reading. The status-page lane is OpenAI's own declaration, shown as they made it.
Third-party chatter, labeled
Public posts from the last hour that mention OpenAI next to an outage word (“down”, “outage”, “not working”). This is what other people said in public, collected by keyword: not something we measured, and not the user reports above. It never affects the verdict.
- Bluesky posts1 in the last hour (prior day averaged 0.8/hour)
- Confirmation probe22:51 UTC, Canada: a new incident on their status page prompted an extra probe. Result: OpenAI answered normally.
- Confirmation probe22:51 UTC, US-East: a new incident on their status page prompted an extra probe. Result: OpenAI answered normally.
- Confirmation probe22:51 UTC, Europe: a new incident on their status page prompted an extra probe. Result: OpenAI answered normally.
- Confirmation probe22:51 UTC, Asia-Pacific: a new incident on their status page prompted an extra probe. Result: OpenAI answered normally.
- Confirmation probe22:50 UTC, US-West: a new incident on their status page prompted an extra probe. Result: OpenAI answered normally.
Checks passing by region
Checks passing is the checks that passed, divided by every check we attempted, per region. Not interpolated, not smoothed, and never computed for a region with zero recorded samples.
| Region | Checks passing | Probes recorded |
|---|---|---|
| US-East | 100.00% | 52739 |
| US-West | 99.99% | 51678 |
| Europe | 99.98% | 49778 |
| Asia-Pacific | 99.86% | 45406 |
| US-Central | 100.00% | 22975 |
| Canada | 99.99% | 22989 |
| UK | 99.99% | 22910 |
| Southeast Asia | 99.99% | 22505 |
| Australia | 99.92% | 22423 |
| South America | 99.73% | 22729 |
Their declared record
OpenAI's own incident history, taken from their public status feed. This is their account of what happened, not our check reading.
Showing the 10 most recent of 114 incidents OpenAI has declared in the history we hold.
These entries are OpenAI's own declarations, taken from their public status feed, shown separately from our own check readings. Their page publishes what they choose to publish: an incident missing here means they did not declare one, not that nothing happened, and an incident here is their account, not our measurement. Neither record ever changes the other.
OpenAI's own status feed works, so its declared incidents are the record and our probe events are not listed or counted. Our check results by region are above, as a reading beside the statement.
Full detail on what a probe does and does not prove is in the methodology page.