Market

Routing and upstreams

For each request, the router picks the cheapest healthy upstream that serves the model you asked for, and falls back if it fails.

Upstreams

ORBIX buys inference through official provider APIs. At launch:

  • OpenRouter for broad model coverage.
  • Direct providers (for example DeepSeek) where the direct price is lower.

New upstreams are added per model when they are cheaper or more reliable. The model you request never changes; only where it runs.

Selection

For every request the router:

  1. Lists upstreams that serve the requested model id.
  2. Drops any that are failing health checks or over their error budget.
  3. Picks the lowest effective price, breaking ties by recent latency.

Fallback

If the chosen upstream returns a 5xx, times out or rate-limits us, the router retries the next upstream once. For streaming requests, fallback only happens before the first token is sent.

Billing follows the upstream

Token counts come from the upstream's own usage report, so your bill matches what the provider measured. The list price used for your discount is the market list price for the model, not the specific upstream's rate.