A token is a sub-word unit produced by the model's tokenizer, roughly 3 to 4 characters or about 0.75 words in English, so 1000 tokens is very roughly 750 words.
Cost and latency both scale with token count because the model does work proportional to the number of tokens in the input plus the number it generates, and providers bill separately for input and output tokens, with output usually more expensive. Tokenization is also why context windows are measured in tokens, why non-English text and code can be more expensive (they tokenize less efficiently), and why you cannot reliably 'trim to N characters' to fit a budget. Practically you should measure real token counts with the provider's tokenizer rather than guessing, because a prompt that looks small can blow your budget once you stuff in retrieved documents or long histories.
- Assuming one token equals one word.
- Ignoring tokens when estimating cost and limits.