The basic unit of text a language model reads and writes, roughly a word fragment.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
A language model never sees words or characters the way a person does. Text is first split into tokens: a common word may be one token, a rarer word several pieces, on average about three to four English characters each. Everything downstream is measured and priced in tokens: context windows, API bills, generation speed.
When a model "writes," it is choosing one next token at a time from a probability distribution over its whole vocabulary. The tokenizer's choices quietly shape behavior, arithmetic, spelling, and non-English text are all harder where the unit boundaries fall awkwardly.