What one agent loop actually costs.

Each step resends the transcript. Input tokens pile up faster than the step count, unless the prefix is cached. Looptally prices that shape with published list rates from OpenAI, Anthropic, and Gemini.

Model

Stable prefix tokens

System prompt and tool schemas. Sent on every step.

New tokens each step

Tool results and new instructions. Not the model’s own output.

Output tokens each step

Include reasoning tokens if you want a number shaped like the bill. Looptally cannot see hidden reasoning.

Steps in the run Runs per day Hours the cache is stored

Used only when the vendor charges cache storage by the hour.

Cache the growing prefix Batch rates

Cache write window

Shown when the vendor publishes two write prices. OpenAI’s table has one.

The numbers in the form are an example, not a benchmark. Replace them. The method page shows the recurrence you can check by hand. The estimate page turns pasted text into a rough token count.

Loading the price book…

Same loop, every model in the book

What the receipt is willing to claim

List prices retrieved 2026-09-27, linked on the price page.
A growing transcript: step k sends the stable prefix, k chunks of new tokens, and the outputs of the earlier steps.
Cache writes, cache reads, and Gemini storage, only when that price is in the book.

It does not know your retries, your hidden system wrapper, your reasoning tokens, or whether the provider actually cached a short prefix. Those gaps are listed on the method page. One filled-in case is on the worked example. For a file you can edit when a vendor changes a price, there is a $19 workbook.