Cache hit rate dominates your total cost calculation for long agent session, and it largely depends on the provider. Deepseek's native deployment is probably much better than third party in this regard. For v4 pro it's a whopping >100x price difference between normal input vs. cached input tokens.