OpenAI
Answering100%
answered, 24h
100%
answered, 7d
1.5s
median reply
143 of 143 checks measured in the last 24h
Online… visits today
Provider status
Provider status checked every 10 minutes
We ask every provider one small question every 10 minutes and record what comes back. This is that record, kept for 7 days.
- 6/6
- providers answered the last check
- 10 min
- between checks, the same call for every provider
- 7 d
- of record kept, with the checks we could not make listed separately
Last 7 days
Every provider, hour by hour
Each bar is one hour of checks. A check that failed on our side is hatched and counts as neither uptime nor downtime, and an hour with no check at all stays empty, so a gap in the record shows as a gap.
All 6 providers answering · last check —Next check in --:--
ProviderNowLatencyLast 7 days, hourlyUptime
OpenAIAnswering1.5 s100%
AnthropicAnswering1.4 s100%
GoogleAnswering629 ms100%
DeepSeekAnswering992 ms100%
Moonshot (Kimi)Answering1.3 s100%
Z.ai (GLM)Answering1.7 s100%
AnsweredSlowDid not answerNot measured — our side, counted as neitherNo check
By provider
The last 24 hours and 7 days
“Answered” counts only the checks we could make. The median reply is our round trip, from a machine that is often running a benchmark sweep at the same time.
Anthropic
Answering100%
answered, 24h
100%
answered, 7d
1.4s
median reply
143 of 143 checks measured in the last 24h
100%
answered, 24h
100%
answered, 7d
0.6s
median reply
143 of 143 checks measured in the last 24h
DeepSeek
Answering100%
answered, 24h
100%
answered, 7d
1.0s
median reply
143 of 143 checks measured in the last 24h
Moonshot (Kimi)
Answering100%
answered, 24h
100%
answered, 7d
1.3s
median reply
143 of 143 checks measured in the last 24h
Z.ai (GLM)
Answering100%
answered, 24h
100%
answered, 7d
1.7s
median reply
143 of 143 checks measured in the last 24h
The record
What we recorded
Outages we recorded
No provider failed to answer in the last 7 days.
Checks we could not make
These were issues on our side, not the providers’. We list them so that a problem of ours never shows up as somebody else’s downtime, and because a gap in the record is a fact about the record. Each line says how long it lasted and when it last happened: anything not marked ongoing is over, and stays listed only until it falls out of the 7-day window. This list starts from 11 September 2026. Earlier checks are still recorded and still count in the uptime figures.
- Anthropiclasted 30 minan issue on our side · 1 episode · last seen 5 days ago, since resolvedAn issue on our side. We are working to resolve it.
Slow periods
Two or more consecutive checks that took longer than 5 seconds. Measured from our machine, which is often running a benchmark sweep at the same time, so read these next to the median above rather than on their own.
- Moonshot (Kimi)lasted 8h10 episodes in the last 7 days
Method
How this is measured
- The check
- One short completion on our own key, from one machine in Germany. It is the same call for every provider, so the numbers are comparable.
- What “answered” means
- The provider returned usable content. It does not mean the model was correct, fast, or good — that is what the benchmarks are for.
- When a check fails
- We ask again fifteen seconds later and record the second answer, so one dropped connection is not published as ten minutes of outage. If the provider still cannot be reached, we test our own side before blaming theirs: whether our server stalled while the check was in flight, whether other providers failed to connect at the same moment, and whether two independent control endpoints answer. Any of those means the failure was ours — it is listed under checks we could not make and counts as neither uptime nor downtime.
- What we do not claim
- One probe, every 10 minutes, from one machine in Germany, on one account. A provider can be down for us and up for you, or the reverse. If your own traffic disagrees with this page, trust your own traffic.
- When a provider is down
- We stop benchmarking it rather than publishing a number we did not measure, and run the missed models as soon as it answers again. An outage leaves a gap in the charts — that gap is deliberate.
Generated 2026-10-23 22:13 UTC · refreshed every minute