Currently result.context_wrapper.usage aggregates all token counts across requests. I'd like to be able to track the individual underlying requests' usage, as for some models (Sonnet 4.5) the price-per-token in a single request differs based on their count, and the cost of a turn can't be calculated easily.