A token is the unit a language model actually reads and writes — usually a common word, a word fragment, or a piece of punctuation. As a rough rule for English, one token is about four characters, and 100 tokens is about 75 words.
Also called: tokenization · Reviewed
Before a model sees your text it is split by a tokeniser into pieces drawn from a fixed vocabulary. Frequent words like "the" are a single token. Rarer ones get broken up — "tokenisation" might become three. Numbers, code and emoji fragment more aggressively than prose, and languages that were less represented in training data often need two or three times as many tokens to say the same thing.
Tokens are the unit of both price and capacity. API pricing is per million tokens, split between input and output; context windows are measured in tokens; rate limits are tokens per minute. Getting a feel for the count is the difference between a prototype that costs a few rupees a day and one that quietly costs thousands.
It also explains behaviour that otherwise looks like a bug. A model that miscounts the letters in a word is not being careless — it never saw the letters, only the tokens the word was split into.
Estimating tokens by counting words is fine for a back-of-envelope figure and wrong wherever it matters. Use the tokeniser for the specific model you are calling: vocabularies differ between model families, so the same prompt genuinely costs different amounts on different providers.
The other trap is forgetting that output tokens are usually several times the price of input tokens. Asking for a shorter answer is often a bigger saving than shortening the prompt.