The unit price of running AI in production.
Providers bill per million tokens, quoted separately for input and output. It is the number that decides whether a product's unit economics work.
Prices for a given capability level have fallen steeply and repeatedly, while frontier prices stay high — so the same feature gets cheaper each year if you are willing to move down a model tier.
Estimating a real bill means multiplying by requests, average prompt length and, critically, the output length you allow.