Cached input tokens per dollar

Beside the input price and the output price, most listings this site reads carry a third: a rate for cached input. It applies to the part of a prompt the host processed in a recent request and can reuse.

That rate is below the normal input price. Turned over, it gives a third tokens-per-dollar figure, higher than the one for fresh input.

Only a repeated opening qualifies. Requests must begin with the same tokens, and arrive within a short time of each other. A long system prompt, a fixed set of tool descriptions and a reference document sent with every question are the usual cases.

Everything after the first changed token is fresh input, billed at the normal rate.

Split your input in two: the share that repeats and the share that does not. Price the first at the cached rate and the second at the input price. Add output at the output price. Divide your tokens by that cost.

For an agent that resends a long history at every step, most of the input repeats. Its real tokens per dollar is well above what the plain input price suggests.

Order the prompt from fixed to changing. Instructions and reference text first, the user’s question last. Do not put a date, a name or a request id at the top.

Then check how the host caches. Some do it for every request. Some need the request to mark what to cache, and some charge for writing an entry. A cached rate you have not switched on buys nothing.

Runcrate charges input and output for all 124 models to one prepaid balance. Compare a week before and a week after you reorder a prompt, and the balance shows what changed. Its live input and output prices are in the tables on this site. Runcrate lists GLM 5.3 at $0.59 per million input 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.