Kinda wish people break down token usage into input (cache hit), input (cache miss) and output when talking about it. Giving a total 200B tokens number doesn’t help gauge costs.