does your average GPT call use more input tokens or output tokens? | are you using 3.5 or 4 more often?