One model name. Every model behind it.
Point your client at orcarouter/auto and every prompt is graded and sent to the model that clears your quality bar at the lowest price — with semantic embedding as an opt-in upgrade.
- Four objectives per workspace — Cheapest, Balanced, Quality, or Adaptive, which learns the trade-off from your own traffic.
- Online learning: the router keeps updating from live production outcomes, not a frozen benchmark run.
- Routing adds under 1ms by default; opt into semantic embedding and you also pay for the embedding call. Either way the grade and chosen model land on every receipt.
40+ providers, one endpoint
The same OpenAI-compatible request reaches any provider, and health-aware balancing spreads load across the ones that are up. Per-channel health, weight and priority take unhealthy upstreams out of rotation automatically, and provider-specific quirks are normalised in the adapter so your client code never branches.
One name. Every model.
You write orcarouter/auto once. From then on, each prompt is read on its way through and sent to whichever model answers it best for the least money. A new frontier model lands on a Tuesday and your traffic starts using it — no migration, no client release, no meeting.
A default, not a black box.
Point a workspace at the objective you actually want. Cheapest. Quality. Balanced. Or Adaptive, which learns the trade-off from your own traffic instead of someone else’s benchmark. Every choice it makes arrives on the receipt with the grade that produced it.