The detail
Tokens are the units AI models actually read and bill for, and they do not line up one-to-one with words. A token is a chunk of characters produced by the model’s tokenizer: frequent whole words like "the" or "cat" are one token, but a longer word such as "tokenization" can be split into several pieces ("token", "ization").
Across normal English prose the long-run average works out to roughly 1.33 tokens per word, or equivalently about 0.75 words per token. That ratio is what most cost and context estimates are built on.
The ratio rises for text that is not plain English: source code, JSON, URLs, mathematical notation, emoji, and non-Latin scripts all tokenize more heavily — sometimes one token per character or less. For exact numbers on real text, paste it into the token counter and pick your model.