Running a model isn't binary, it's per amount of time spent generating tokens.