Loading status…
Replicate status · AI & APIs · checked (43s ago)
We could not reach Replicate's status feed, so its current state is unconfirmed.
According to Replicate's official status page, Replicate was reporting “Status API unreachable (Unexpected token '<', "<!DOCTYPE "... is not valid JSON)” when isuptime last checked at . Over the last 30 days, Replicate's feed reported no outage in 100.00% of our checks.
isuptime cannot fix Replicate, and neither can you. These are the things that are actually worth doing until it comes back.
Everything above is read from Replicate's own status feed. This is what other visitors are saying in the last 24 hours — useful when the feed is green and you are still stuck, and never used to change the status above.
Nobody has reported Replicate in the last 24 hours. Yours would be the first.
Reports are anonymous counts and are kept for 24 hours, then deleted. No account, no message, nothing stored about who reported. One report per service per hour, changeable until the hour is out.
Replicate runs machine learning models as hosted APIs — image generation, video, audio, language models and anything else pushed as a Cog container. You call a model, it returns a prediction, and Replicate handles the GPUs underneath.
The status page covers the API and the prediction pipeline. The distinctive failure mode is not covered by any indicator: models scale to zero when idle, so the first call after a quiet period pays a cold boot that can run to minutes. That is designed behaviour, not an incident, and it is the single most common reason a Replicate call appears to hang.
isuptime reads this feed on a schedule and reports what it says — how that works, in detail.
One bar per day. Each shows the share of that day's checks where Replicate's own status feed reported no outage — degraded counts as no-outage. Days with no bar are days isuptime recorded no checks, not days Replicate was down.
14 other services isuptime tracks were also reporting a problem as of the same check, 43s ago. They are checked on one schedule, so the timing is comparable — but overlapping in time does not mean they share a cause.
Each point is one check of Replicate's status endpoint.
Every incident Replicate has published in the last 12 months.
Replicate has not published a single incident in the last 12 months. That is a statement about Replicate's posting habits as much as its reliability — a status page only records what its owner decides to record.
Check whether the model you are calling is public and still published — authors can remove or rename versions, which fails per-model while Replicate itself is fine. A prediction stuck in `starting` is a cold boot; a prediction stuck in `processing` past the model's usual time is usually the model, not the platform. GPU capacity for a specific hardware tier can be exhausted without a platform-wide incident being posted.
