How the money is counted

Tokens: exact, from the agent’s own log. Money: an estimate at list prices; your bill may differ (plan, contract rates, batch, cloud provider). Cursor and Copilot Chat keep no token counts, so for them tokens are estimated from the size of the text, and the money is shown as a range.

Each kind of token is priced on its own:

Kind	What it is	Price
New input	Text sent to the model for the first time.	Full input price.
Written to the cache	Text the provider stores so it can be re-read cheaply.	Set by the provider; for Claude, a little above the input price.
Re-read from the cache	Earlier chat sent again, served from the cache.	About a tenth of the input price.
Output	What the model writes back.	The output price.

Prices are the providers’ public API prices. The report shows the date of the price list it used.

Your own prices

• For one report: sessioncost --prices my-prices.json.
• Every time: put them in ~/.sessioncost/prices.json.

On a subscription

On a monthly plan (Claude Pro or Max, ChatGPT) you are not billed per token. sessioncost --subscription labels the money as API-equivalent: what the same chat would have cost at API prices. It is a fair way to compare chats, not your bill.