What is token pricing?

Token pricing is how AI providers charge for using their models: by the amount of text going in and coming out, counted in tokens. Prices are usually quoted per million tokens. Text the model writes typically costs around 5 times as much as text sent to it. Many providers give discounts for text that repeats across requests, called caching, and for work that can wait a few hours.

Why it matters for a business

The rate per token is rarely what surprises anyone. The number of tokens is. Most of what gets billed is text nobody typed: standing instructions, pasted documents, records a lookup pulled in, and the whole conversation re-sent on every turn. Reasoning models bill their thinking as output too. An estimate takes 4 numbers that are known before anything is built: runs per month, tokens in per run, tokens out per run, and the model tier. It is most useful read as a cost per run rather than a monthly total.

Example

Let’s say a support tool answers 4,800 order-status questions a month, sending about 1,600 tokens and getting back about 200 each time. At $2 per million tokens in and $10 per million out, that is about $25 a month, half a cent a question. Switching on a reasoning setting that thinks for 1,500 tokens before each reply takes the same month to nearly $100.

Go deeper

How much AI costs, and how token pricing works

Related terms

• Token
• Inference
• Reasoning model
• Context window

All terms, A to Z