deepinfra — AI API latency & reliability by region
As of September 06, 2026, the deepinfra AI API responds in 256 ms p50 (edge, time-to-first-byte) from US (Central) — its fastest measured region — with 100% uptime. Slowest is South America (São Paulo) at 609 ms p50.
Is deepinfra down right now?
Reachable from all 4 regions
Last successful probe Sep 6, 17:15 UTC · every 5 minutes from 4 region(s)
2 isolated failure(s) in the last 72 hours, the most recent (host probes) visible from only one region (South America (São Paulo)). A single-region failure cannot be separated from a problem on that network path, so it is not called an outage. Measured from outside deepinfra’s own network, independently of its status page; full log: measured AI API incidents.
Real requests: not measured for deepinfra — only the API host is probed here, so errors limited to specific models are invisible to these probes. That is exactly what a status page tends to report, hence the line below.
deepinfra publishes no machine-readable status page, so there is nothing to cross-check against.
Edge latency, 24 h
256 ms
p50 from US (Central), the fastest region
Uptime, 24 h
99.9%
1,151 probes across 4 region(s)
Failed probes, 72 h
2
of 3,454 · South America (São Paulo)
Reported by deepinfra, 72 h
—
no machine-readable status page
Was deepinfra 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).
South America (São Paulo) only — not an outage · under one 5-minute probe cycle · 1 failed probe(s) · timeout
South America (São Paulo) only — not an outage · under one 5-minute probe cycle · 1 failed probe(s) · timeout
What deepinfra’s own status page reported
deepinfra publishes no machine-readable status page, so there is nothing to set against the measurements above.
Is deepinfra getting faster or slower?
hourly p50 latencyall probes succeededsome probes failedhalf or more failedno data
One panel per region: the line is the hourly p50 edge latency (shared scale, ms), the strip beneath it is the outcome of that hour’s host probes. Hover any hour for the numbers. Flat is good; a rising line means the endpoint is slowing down from that origin, a gap means no probe succeeded in that hour.
How fast is the deepinfra API from each region?
Region
p50 TTFB
p95
Uptime
Asia (Tokyo)
517 ms
778 ms
100%
Europe (Germany)
600 ms
1097 ms
100%
South America (São Paulo)
609 ms
1437 ms
99.7%
US (Central)
256 ms
918 ms
100%
Measured from US (Central), deepinfra responds in 256 ms; from South America (São Paulo) the same API takes 609 ms — 2.4× 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 deepinfra’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
deepinfra’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 deepinfra’s
own reports are shown next to the measurements rather than folded into them.