The router weights price by its inverse square.
By OpenRouter's published default, providers with a recent outage are skipped and the rest are weighted by 1 / price². One third the price is nine times as likely to be routed to first. Every point of utilization is cost per token you can keep as margin or pass on as price.
Illustration of the published default weighting only: four providers, no recent outages, identical except price. Latency and throughput order providers only when the caller asks for them.
More concurrent tenants and longer contexts before turning traffic away or adding accelerators.
Recovered utilization gives room to price lower on the router, or to keep price and take the margin.
More requests from the same memory keeps capacity rejections from dropping you when demand peaks.
API contract, model versions and fleet topology stay as your router listing depends on them.
Find out whether memory is costing your fleet.
The first step puts nothing in your request path: a couple of weeks of counters your serving engine already keeps, and a written answer. Then we confirm it on your own traffic, on your own hardware, against your current configuration.