I’m looking to better understand Anthropic’s prompt caching functionality for the Claude API:

1. How should ephemeral properties and demo caching be handled after
   initial caching? do I have to keep calling the cache property or remove it?
2. With multiple chatbots that have different system prompts but are sharing an API key, is caching shared or isolated between them?
3. Does cached content transfer when switching between Claude models (e.g., from Haiku to Sonnet), or does it need to be re-cached?

Any insights on these implementation details would be helpful.