How do you handle cached input or long-context pricing? Some providers change the rate a lot depending on that, so I’m curious how accurate the estimate stays.