Definition
A token is the basic unit of text that a language model reads and generates. A tokenizer splits raw text into tokens; a token may be a word, a subword, a character, or a punctuation mark or other symbol.
How it works
The model does not operate on raw characters — each token is mapped to a vector that participates in the computation, and both input and output are sequences of tokens. The number of tokens a piece of text consumes drives compute, cost, and the usable context length.
Usage
Developers watch token counts when managing cost, sizing requests, and planning context capacity. Usage-billed APIs and the progress of streaming output are likewise measured in tokens.
Related and confusable terms
The model emits one next token at a time, at a rate expressed in tokens per second; how many tokens a request can hold is bounded by the context window. All of this happens inside a large language model.