For tuning and model selection purposes it'd be nice to see how many tokens an API call used for input and output. And on top of that show how much that API call cost.