💭 is there any way to make a long llm chat with text message length messages cost effective? I think even the most generous caching wouldn't help