Sounds reasonable until you consider that the "prompt" might include a million tokens of context, not to mention follow-up/iterations