How routing and failover work

Each model has a ranked list of upstream providers (see Models). The ranking rules:

1. Verified full precision first. Providers serving the maker’s weights unchanged come before providers that don’t state their precision. Quantized copies are never used.
2. Cheapest first within each tier.

Failover: if a provider fails (error, rate limit, outage), TOKEN retries your request on the next provider, up to three times. The price you pay is the model’s listed price, whichever provider answers.

Health checks:

• A provider that keeps failing is removed from rotation automatically.
• It is re-added when its health checks pass again.
• Per-model availability is published on the status page.

Single provider: some models (labelled single provider) have only one upstream today. They have no failover until a second provider is available; we add one automatically when it appears.

What TOKEN stores: token counts and request metadata, never your prompts or completions. See the privacy policy.