Guides

Failover

Each model can have several routes. If one fails, we try the next healthy one.

How a route is picked

  • Only healthy routes are used: above 90% success recently, and not tripped by repeated errors.
  • If your workspace needs zero data retention, only ZDR routes are considered.
  • Among those, the route that costs us least is tried first.

Retries

  • Up to three tries per call.
  • Content policy refusals are never retried on another route.
  • For text, fallback only happens before the first token.

What you pay

The price book price for the call that worked, whichever route served it. Failed tries cost you nothing. If every route fails, the hold is released in full.