Full definition
Before replying, a language model splits text into tokens, word fragments of around 4 characters (approximately 0.75 words). ‘chat’ is one token, ‘anticonstitutionnellement’ is several. It is the native API unit: billing and energy measurement are per million tokens. A request’s energy follows the number of tokens processed, especially output tokens: a long prompt or response consumes more. Order of magnitude: around 3x10⁻⁴ Wh per token on an optimised model.