Term: Token pricing
~6 min read
Estimated time: ~6 min read — for the in-app brief plus opening the primary source.
What this is
Most APIs charge separately for tokens in (what you send) and tokens out (what the model writes). That is the unit cost of AI work.
Everyday example
A vendor quotes $3 per million input tokens and $15 per million output tokens. A weekly board-pack summary that sends 40 pages in and writes 2 pages back has a calculable unit cost — that is token pricing.
Token pricing is the rate card: you pay for tokens in and, usually more, for tokens out.
- Seat licenses hide this; APIs do not.
- A cheap model on a huge paste can beat a dear model on a tight brief — or the reverse.
- Retries, long chats, and verbose answers multiply the bill.
- Consumer ChatGPT/Claude/Grok plans are often bundled; enterprise APIs are typically metered.
Next action: Before the next pilot, estimate tokens per task × volume × list price, plus a retry buffer.
What changes in how you lead
How decision rights, process, and ownership should change.
- Finance sees a usage forecast, not only a per-seat line.
- Purchasing writes a spend cap and an overage conversation into the order.
Compare related ideas
Token pricing vs Token
A token is the unit of text. Token pricing is the price list applied to those units.
Open TokenToken pricing vs Training vs inference
Your day-to-day bill is almost always inference (using the model), priced in tokens — not the lab’s cost to train the foundation model.
Open Training vs inferenceDeep dive
Input is usually cheaper than output. Long answers cost more than long questions.
Purchasing: model price × expected monthly tokens for the process — plus a buffer for retries and chatty users.
Finance: cost per successful ticket resolved or memo produced, not cost per login.
Marketing: cost per asset variant that passes review.
Consumer ChatGPT / Claude / Grok subscriptions are a different commercial model (flat or bundled). Enterprise APIs are typically token-priced — read the rate card.