Anthropic prompt caching — worth the setup for a chatbot?

We're paying a lot for repeated system prompts across a long agent conversation. Their caching docs promise 90% savings on reads. Anyone measured real-world savings? Does it break with streaming?