Tokens

The chunks (typically 1-4 characters or ~0.75 word) that LLMs read and generate. Model input/output limits, pricing, and context windows are all measured in tokens.