Your input to output mix

Every API response reports how many input tokens and how many output tokens the request used. Log both for a sample of real traffic. Divide total input by total output. That ratio is your mix.

Do it per workload, not for the whole account. A support bot and a nightly summarizing job on the same key have very different mixes.

Most production work is input-heavy. Answering from retrieved documents sends pages of text for a paragraph back. Classifying, extracting fields and summarizing send far more than they return. Agents resend a growing history at every step.

For this work the input price carries the bill, and the output price hardly shows. A cached input rate, where the host has one, matters more than either.

Writing from a short brief, generating code and translating return as much as they receive, or more. Here the output price carries the bill.

Reasoning models shift any task toward output, because their thinking is billed as output tokens.

Two listings of one model often split the lead. One has the lower input price and the other the lower output price. Which is lower for you depends on the mix.

A blended price is a shorthand: one figure for a fixed mix, commonly three parts input to one part output. It ranks fairly for work near that mix. For anything else, do the sum with your own.

Runcrate puts 124 models behind one OpenAI-compatible API and one prepaid balance. You can send a summarizing job to a model with a low input rate and a writing job to one with a low output rate, from the same key. Runcrate lists MiniMax M3.1 Flash at $0.158 per million input tokens and $0.63 per million output tokens.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.