A token is the smallest unit of text a model processes. Tokenizers split text into common words, word fragments, punctuation, and bytes, so "unbelievable" may be three tokens and "the" one. English prose averages roughly four characters per token, but code, JSON, Chinese, and emoji all shift the ratio enough that estimates from word count are unreliable.
Tokens are the unit for three separate limits: the context window, the price, and the rate limit. A request's input tokens include the system prompt, the conversation, retrieved documents, and tool schemas; output tokens include the answer and, on some models, hidden reasoning that is billed but not shown.
When a budget matters, count with the vendor's tokenizer rather than guessing. The token cost guide shows the arithmetic, and the context window entry explains why the limit is shared by input and output.