Arbiter is a drop-in OpenAI-compatible router. It sends each request to the best model-for-the-money on the BTL runtime, checks the answer, and learns, so you stop paying premium prices for simple work.
Saved vs baseline
Calls routed
Models in pool
5
Task types
How it works
Change one base URL and keep your OpenAI code exactly as it is. The model field you send is ignored - choosing the model is Arbiter's whole job.
Arbiter classifies the request, drops any model whose context cannot hold it, and picks the cheapest one that is still good enough for that kind of task.
Every answer is graded and priced from the runtime's own cost headers. Savings are measured, not guessed, and climb as the router learns.
Why the runtime
One key reaches many models, and every response reports its real cost. A raw provider key gives you neither. That is exactly why the measured savings and the price-reaction can only exist on top of the runtime - Arbiter routes model choice, the runtime routes providers and reports the bill.
Send a real request and watch it get classified, routed, scored and priced live.