The API does tell you how many prompt and completion tokens each request used, if you're okay knowing after-the-fact