Tokens are the actual unit of currency for language models — not words, not characters. Model pricing, context windows, and rate limits are all measured in tokens. One English word averages roughly 1.3 tokens. Code, non-English text, and unusual words use more.