What is a token

Every price on this site is per million tokens, and every context limit is counted in them. A token is the unit a model actually reads and writes: a short chunk of text, usually a word or part of one. As a rule of thumb in English, a token is about three-quarters of a word — this sentence is around twenty tokens.

Why models bill by it. A model's work is proportional to the text it processes, and tokens are how that text is measured. Nothing else about your request matters to the bill: not time, not "questions", not sessions. Tokens in, tokens out.

Input and output are priced separately, and the gap is large — output typically costs several times more per token, because generating text is more work than reading it. This has a practical consequence people miss: for many real workloads input still dominates the bill, because every request re-sends instructions and context. A question of ten words on top of a two-page system prompt is not a ten-word request. The model pages always show both prices; a price pair always comes from one host.

Rough intuitions worth keeping. A paragraph is on the order of a hundred tokens; a long email, several hundred; a short report, a few thousand; a novel, a few hundred thousand. So a million tokens — the unit prices are quoted in — is genuinely a lot of text, and this is why metered API use often costs less than people fear.

One honest wrinkle. Tokenizers are tuned on the text they were trained on, which skews English-ward: the same sentence in Danish, Japanese or Hindi often needs noticeably more tokens than its English translation. Same meaning, bigger bill, smaller effective context window. If you work mostly outside English, budget for it.

Where next: What is a context window · Leaving a flat-rate subscription · Model prices