How do you handle prompt caching? A lot of cost savings for a single model chat come from cache hits on the conversation context, and switching models invalidates that cache — the new model has to reprocess everything at full input price.