What is the openai API's uptime?

Measured uptime for the openai API averages 99.8% across 4 region(s) over the last 24 hours.

Is the OpenAI API down? Live status, latency & uptime by region

Last verified October 21, 2026 · rebuilt about every 30 minutes

Is the OpenAI API down right now?

Reachable from all 4 regions, requests to gpt-4o-mini succeeding

Last successful probe Oct 21, 20:16 UTC · every 5 minutes from 4 region(s) · status feed checked Oct 21, 20:17 UTC

2 isolated failure(s) in the last 72 hours, the most recent (host probes) visible from only one region (Europe (Germany)). A single-region failure cannot be separated from a problem on that network path, so it is not called an outage. Measured from outside openai’s own network, independently of its status page; full log: measured AI API incidents.

Real requests to gpt-4o-mini: 0 of 1 succeeded in the last 24 hours from 1 region(s); last failure Oct 21, 10:26 UTC from Europe (Germany) ([Errno 101] Network is unreachable). A one-token completion is sent on the same schedule as the host probe; 5xx, 529 and timeouts count as failures, problems with our own key or quota are excluded.

OpenAI’s status page lists no open incident (checked Oct 21, 20:17 UTC); its 4 entries from the last 72 hours are listed below.

As of October 21, 2026, the openai AI API responds in 148 ms p50 (edge, time-to-first-byte) from US (Central) — its fastest measured region — with 100% uptime. Slowest is Europe (Germany) at 294 ms p50.

Endpoint response (TTFB), 24 h

148 ms

Typical (p50) from US (Central), the fastest region

Uptime, 24 h

99.8%

1,152 probes across 4 region(s)

Failed probes, 72 h

of 3,459 · Europe (Germany)

Failed requests, 24 h

of 1 to gpt-4o-mini

Reported by OpenAI, 72 h

on status.openai.com

Was OpenAI down in the last 3 days?

No. 2 isolated single-region probe failure(s) were recorded, none of which meets the bar for an outage (two or more regions at once). OpenAI’s own status page reported 4 incident(s) in the same window, listed below.

Oct 21, 10:36 UTCEurope (Germany) only — not an outage · at least 5 minutes · 2 failed probe(s) · [Errno 101] Network is unreachable
Oct 21, 10:26 UTCrequests to gpt-4o-mini: Europe (Germany) only — not an outage · under one 5-minute probe cycle · 1 failed probe(s) · [Errno 101] Network is unreachable

Earlier days with OpenAI incidents: Oct 19 (status page) · Oct 15 (status page) · Oct 14 (status page) · Oct 13 (status page) · Oct 9 (status page) · Oct 8 (status page) · Oct 7 (status page) · Oct 6 (status page)

Show this live in your README: — embed code

What OpenAI’s own status page reported

Oct 21, 15:34 UTCminor Elevated error rates for gpt-5.6-luna API requests · resolved Oct 21, 16:44 UTC
Oct 21, 14:17 UTCminor Elevated FedRAMP API error rates for Astra and Sol · resolved Oct 21, 19:34 UTC
Oct 21, 02:26 UTCminor Degraded performance in ChatGPT, Codex, and APIs · resolved Oct 21, 03:28 UTC
Oct 19, 13:31 UTCminor File Download errors · resolved Oct 19, 14:14 UTC

Self-reported by OpenAI, checked Oct 21, 20:17 UTC; not verified by our probes. Errors on individual models are visible only from inside the provider.

Is OpenAI getting faster or slower?

Endpoint response (TTFB) · milliseconds · lower is faster

Last 72 hours · UTC · newest time on the right · current hour is incomplete.

Each line shows a region’s typical response per hour (p50), using the same scale. A gap means no successful timed probe in that hour.

typical response per hour (p50)all probes succeededsome failed (under half)half or more failedno probes recorded

Probe counts by region

Europe (Germany): 2 failed / 858 endpoint probesEurope (Germany) · model: 1 failed / 1 requests · 71 hour(s) without dataUS (Central): 0 failed / 856 endpoint probesAsia (Tokyo): 0 failed / 856 endpoint probesSouth America (São Paulo): 0 failed / 856 endpoint probes

Chart details

The host strip shows endpoint probe outcomes; the model strip shows completion-request outcomes for gpt-4o-mini. The model strip does not show time to first token (TTFT). Adjacent hours with all successes or no data are joined. Hover an hour for its recorded values.

Latest is the most recent hourly p50 with data. Median is the median of hourly p50 values over these 72 hours. The “slower than usual” label means latest is at least 25% above that median; it is a relative change, not a fixed speed rating.

How fast is the OpenAI API from each region?

Endpoint response (TTFB) · last 24 hours · milliseconds (ms) · lower is faster.

How to read these metrics

Endpoint response (TTFB)Network connection and the first response byte, without model inference.

Typical · p5050% of successful timed samples finish within this time.

Slower · p9595% of successful timed samples finish within this time.

Region	Typical (p50 TTFB)	Slower (p95)	Uptime
Asia (Tokyo)	216 ms	302 ms	100%
Europe (Germany)	294 ms	478 ms	99.3%
South America (São Paulo)	264 ms	544 ms	100%
US (Central)	148 ms	233 ms	100%

Measured from US (Central), openai responds in 148 ms; from Europe (Germany) the same API takes 294 ms — 2.0× longer for an identical request. That difference is the network path, not the model, and it applies to every call you make.

What is actually being measured here?

A probe in each region opens a real connection to openai’s own API host every five minutes and times DNS resolution, the TCP handshake, the TLS handshake and the first byte of the response. Nothing is routed through a gateway or aggregator, so these figures describe openai’s infrastructure rather than a reseller’s. Uptime counts a probe as failed only when the service genuinely fails — a 401 from an unauthenticated probe means the endpoint answered correctly and is counted as up.

What this page does not tell you

These are edge latency numbers: the time before the model begins generating. They say nothing about answer quality, throughput, price, or how long a full completion takes — those depend on the model you call. Edge latency matters because it is unavoidable: it is paid on every request, before a single token exists, and no prompt engineering removes it. Where an API key is available, inference time-to-first-token is measured separately and shown on the region pages with the model named next to the figure. Likewise, an unauthenticated probe cannot see errors that hit only some models or some accounts — the kind a status page usually reports — which is why OpenAI’s own reports are shown next to the measurements rather than folded into them.