Token
The unit a language model uses to read and write text — roughly three-quarters of a word on average, though common words are one token and rare ones may be several.
A token is the atomic unit a language model works in. Text gets broken up ("tokenized") before the model sees it, and the model outputs one token at a time.
Rough rule: one token is about 0.75 words on average in English. Common words ("the", "and") are one token. Uncommon or long words ("neuroscience", "antidisestablishmentarianism") get split into multiple tokens. Punctuation, whitespace, and formatting are their own tokens. Non-English languages tokenize differently, sometimes with worse compression — Japanese and Chinese often use more tokens per unit of meaning than English.
Tokens matter for two practical reasons. First, pricing: model APIs charge per input and output token, so a chatty response literally costs more. Second, context windows: the 200K context window in Claude Sonnet means 200K tokens of combined input and output, which sets an upper bound on how much material fits in one conversation (~300 pages of English prose).