Smallest unit of text a model processes, roughly a word piece; context limits are measured in tokens and billed per token in many APIs.