Skip to main content
When several providers serve the same model, default routing ranks them by expected cost per successful answer and shares traffic by that ranking: the lowest cost wins the largest share. This page is for providers who want to understand that ranking and win more of it.

​
What the ranking measures

For each endpoint, Phaseo measures:
  • Effective price. What a typical request for the model costs on your endpoint, using the model’s real prompt and output sizes and the share of prompt tokens you actually serve from cache. A low input price does not help if output or cache reads cost more.
  • Time to first token, counted from when Phaseo sends the request.
  • Output speed, measured across the whole response, first token included.
  • Errors you cause, such as 5xx responses and streams that fail partway. These carry a heavy penalty.
  • Rate-limit rejections (429). These carry a smaller penalty.
Waiting time is converted to dollars at the model’s typical price per second, so a faster endpoint can charge more and still rank first.

​
Try it

The sliders set your endpoint against three other providers of the same model. The curve shows how traffic falls away as your cost per answer rises above the lowest.

​
How traffic is shared

  • Providers within about 30% of the lowest cost per answer share most of the traffic. Being slightly behind costs only slightly.
  • Beyond that band, share falls quickly.
  • Every routable endpoint keeps a small share, so recoveries and improvements are noticed without any action from you.
  • A conversation that is reading from your prompt cache usually stays on your endpoint while it remains healthy.
Requests that name a provider order, or ask to sort by price, latency or throughput, do not use this ranking. See Routing and fallbacks.

​
What earns more traffic

  • Lower the price of a typical request, including output and cache reads.
  • Serve repeated prompts from cache.
  • Return the first token sooner and stream faster.
  • Avoid errors and rate-limit rejections, and add capacity before demand reaches your limits.
Last modified on October 30, 2026