Lexicon · II · Access
Token
Also: text token
The model's unit of account: a fragment of text worth about three quarters of a word on average. Everything is measured in tokens: document length, limits, and the bill.
The analogy
The taximeter. A ride is billed neither per trip nor per full minute: it is sliced into units that tick by. The model does the same with text: every fragment read and every fragment produced makes the meter run. Understanding tokens is understanding why "ask in one line" and "paste in the 400 pages" do not cost the same.
A concrete example
"Hello, could you summarize this document?" weighs about ten tokens. A forty-page contract weighs around thirty thousand. To budget a use case, you count neither requests nor pages but tokens in and tokens out, which is exactly how providers publish their prices, per million tokens.
What it is NOT
It is neither a word, nor a character, nor a syllable: the slicing is statistical, not grammatical. It is not a unit of meaning either: the model does not "understand" token by token, it processes the whole. And it is not a technical footnote: context limits, latency and costs are all reasoned in this unit.
Often confused with
The distinction that changes the conversation.
- Word
- The model does not read words: it slices text into fragments. 'Antidisestablishmentarianism' is several tokens, 'of' is one. Quotes, quotas and limits in AI are counted in tokens; converting them to 'words' in your head skews the orders of magnitude by roughly a third.
- Access token (API token)
- Same word, different world: an access token is a security key proving an identity to a service. The model's token is a unit of text. In an AI meeting, both come up; always specify which one you mean.