Status
Measured from the requests we actually served, not from a prober, and not from a badge somebody else hosts. The slow requests are in these numbers. That is the point of publishing them.
All times UTC · read at most 60s before you loaded this
Time to first token, last 30 days
p95 first, deliberately. The median is the number that flatters; the 95th is the one that decides whether your agent feels stuck.
Not enough measured days yet to draw a trend. The numbers above are the whole of what has been measured.
Nothing warms up and nothing scales to zero: every model in the catalog is served by an upstream provider that is already running. There is no cold-start penalty to disclose because there is no cold start. What you can be charged by instead is upstream queueing under load, and that is already inside the p95 above; it is not filtered out.
We do not throttle by tier. Concurrency limits are published per tier and returned on every response in X-Tium-Concurrency-Limit, so the limit you are subject to is never a hidden one. Your own per-request timings are on Activity, including the slow ones.
Per model
A verdict needs a denominator. Below five requests in the window, this page says it does not know rather than turning one upstream hiccup into a red dot.
No requests in the last hour, so there is nothing measured to report. This is not a claim that it is down.
- p50 TTFT
- —
- p95 TTFT
- —
- last hour
- 0 req · 0 failed
- last request
- —
No requests in the last hour, so there is nothing measured to report. This is not a claim that it is down.
- p50 TTFT
- —
- p95 TTFT
- —
- last hour
- 0 req · 0 failed
- last request
- —
No requests in the last hour, so there is nothing measured to report. This is not a claim that it is down.
- p50 TTFT
- —
- p95 TTFT
- —
- last hour
- 0 req · 0 failed
- last request
- —
No requests in the last hour, so there is nothing measured to report. This is not a claim that it is down.
- p50 TTFT
- —
- p95 TTFT
- —
- last hour
- 0 req · 0 failed
- last request
- —
Cancelled requests (an agent hanging up mid-stream) are excluded from these counts. Everything else that failed is in them.
90 days of served requests
Share of requests that completed, by day, across every model. A day we served nothing is drawn as a gap, not as a good day.
What went wrong, in plain words
Written by hand, newest first. No severity levels: a taxonomy would tell you less than the sentence would.
Nothing logged yet. This log starts the day the API does; an empty one this early means the service is young, not that it is flawless.