Foundations
Token
The unit models read and write, and the unit you are billed in — roughly ¾ of a word in English.
Models do not see words or characters; they see tokens. A rough English rule of thumb is 100 tokens ≈ 75 words, though code and unusual names use more.
Tokens matter for two practical reasons:
- Cost. You pay per token, usually at different rates for input and output. A feature that quietly stuffs a large document into every request can be far more expensive than expected.
- Limits. Everything you send has to fit the context window.
If you are running AI at any volume, meter token usage per feature and per customer. See cost per call.