AI API status
Every 45 seconds we measure each provider's real-time response from our test node in Chicago (US), timing how long it takes to start answering a short request. Official status pages are checked every five minutes.
How this is measured
Time to first token
Measured from the moment a request is sent until the first piece of the answer arrives. An acknowledgement with no text does not count, because it would make a provider look faster than it is. Connection setup is timed separately.
Uptime
A measurement counts as up when the API answers. Rate-limited answers count as up but show as degraded. Server errors, overload errors and timeouts count as down. Problems on our side, such as an expired key, are left out of the score.
Status combines two signals
A provider shows as degraded or down if our measurements fail or its own status page reports an incident. The worse of the two wins, so an outage the provider has confirmed shows here even if our short requests still get through.
Your numbers will differ
These measurements use small, fast models and very short answers, sent from a US Midwest test node. Larger models, long prompts and other regions are slower. Use "Test your own endpoint" to measure your model with your key.