Model status

How status is calculated

Each bar is one hour of real client traffic, and the unit is a request somebody actually sent: green means requests were answered, amber means a noticeable share came back as an error, red means several failed and not one request came back that hour, grey means there was no traffic to judge by, or an hour we can no longer judge honestly. A single failed request in an otherwise empty hour stays amber — one failure is not an outage. Nor is one caller: when every failure of an hour came from a single client while the model had recently been answering others, the hour stays amber however many requests that client sent, because one client's errors are not yet evidence about the model. A request counts as failed only if the caller received the error: after every account for that model had been tried and every retry spent. What the router had to do to get the answer does not count against the model — rotating to another account, retrying, waiting out a rate limit — because the caller got their answer, and this page publishes their experience rather than ours. A request refused as invalid before it reached any model is not counted either way. Two things to know about the number: a model in a combo that failed while a sibling model answered still counts as failed for itself, since it did not produce that answer; and a reply that broke in the middle after streaming had begun counts as answered, because part of it had already reached the caller. Image and video models are on this page too, and one thing about them is not obvious: a video, and any generation queued in a studio, is recorded as a request the moment the provider accepts the job, so a job that dies later would otherwise stay counted as an answer. When such a generation's charge is returned — the provider produced nothing usable, the job never arrived — that request counts as failed here rather than as served. Refunds that are not the model's fault are not counted against it: a job our own restart or dispatcher interrupted before the provider answered, one that never reached the provider, and an attempt the studio resubmitted, where the retry delivered the result. A generation the provider refused over the caller's own prompt is not counted either way, like a text request rejected as invalid. Hours older than the switch to this method were measured the previous way — by upstream calls rather than client requests — and read lower; they leave the window seven days after the change. The page never exposes account names, keys, request content or internal error details.