Degraded performance

Every minute we check that Console and the inference gateway respond, and look at the outcome of requests over the previous five minutes. For hosted models, that includes streams that fail partway through and responses that are slow to start, even when the request connected successfully. Requests we turn away because of an account’s balance, budget, or usage limits don’t count as failures. Each plan’s status follows its six most-used models over the past week, not counting the free models Go offers; open a plan to see them.

Degraded means a noticeable share of requests are failing, or responses are starting much more slowly than usual for that model. Outage means most requests are failing, or the service isn’t responding at all. A change is confirmed on consecutive checks, so a single failed check appears as degraded. After a service recovers, its incident stays open for a short while in case the problem comes back. Each bar covers three hours and shows the worst status we observed in that time, not an uptime percentage.

Medians over the last five minutes, so half of responses were faster. The first number is the time to the first token, from when a request reaches us; the second is tokens per second while a response streams. A plan’s numbers cover the models it follows; open it to see each model’s. They appear once enough responses were measured. The rate includes reasoning tokens, so it reads higher for models that reason before they start streaming.

By their impact on you. When a model provider has problems, the plans it affects show it here, marked as an upstream provider issue so you can tell it apart from a problem with OpenCode itself. When enough recent requests show it, we name the affected models among those each plan follows, without naming our upstream suppliers.

Yes. GET /api/status returns current status, incidents, and the last week’s history in three-hour periods as JSON. No authentication is required.