Blended price: one number for two prices

Two prices are hard to rank. So comparison tables often combine them. They assume a mix, commonly three input tokens for every output token, and take the weighted average. The result is one price per million tokens that can be sorted.

The mix is an assumption. If your application sends twenty tokens for each one it gets back, a blend at three to one overstates the weight of output. If it writes more than it reads, the blend understates it.

Two models can swap places when the mix changes. One with a low input price and a high output price wins on reading work and loses on writing work.

Take a day of real requests. Add up the input tokens and the output tokens. Multiply each total, in millions, by its price and add the two. Divide by the total millions of tokens. That is your blended price for that model on that host.

If part of your input is cached, split it out and use the cached input price for that part.

Use a blended price to cut a long list down to a short one. Use the two real prices and your own counts to choose from the short list. Any table that ranks by a blend should say which mix it used.

Runcrate lists the input and output price of each of its 125 models, so the blend is yours to compute. Runcrate serves GLM 5.3 Flash at $0.158 per million input tokens and $0.525 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.