Also called: tokens, tokenization
The chunks a model reads and writes - roughly three-quarters of a word each in English.
Why it mattersYou're billed per token and limited by tokens. It's the unit of everything.
See also Context Window, Prompt Caching