Routing & Reliability

Depending on the model, a model id in the catalog can be served by any of several underlying deployments. The platform decides where each request runs — there are no routing parameters to set, and no per-provider configuration on your side. You send a model id; routing, health monitoring, and failover happen automatically.

Responses always echo the model id you requested, and billing always uses that model's published rates, regardless of which deployment served the request.

When a model has more than one deployment, traffic is spread across them according to a weighted split that is continuously adjusted based on observed health — error rates, capacity, and latency. Deployments that degrade shed traffic to healthy ones automatically and regain it gradually as they recover.

Multi-turn conversations are kept on the same deployment where possible, so prompt caching keeps working across requests — send the same conversation prefix and you benefit from cache-read pricing without doing anything special.

When a deployment fails — a provider error, a rate limit, or a timeout — the request is retried on another deployment for that model when one is available. Each attempt is bounded by time-to-first-token and total-duration limits, and exceeding either counts as a failure.

Two behaviors worth designing around:

• Errors you can fix are not retried. Deterministic client errors — an invalid request or a prompt that exceeds the model's context window — surface immediately as a 400 rather than being retried across deployments.
• When every deployment fails, the request returns a 429 indicating the model is overloaded, with a Retry-After header. Retrying after the indicated interval is safe: requests that fail before producing any output are not billed.

Failures before the first token are handled invisibly — failover happens before the stream starts, and you receive a normal response. If a stream fails after output has begun, it ends with an in-band error event; your client should treat an unterminated stream as incomplete. Streamed requests are billed for the tokens actually delivered.