Prompt caching

Providers that support caching apply it automatically to shared prompt prefixes. Cached input tokens are billed at the provider's discounted cache-read rate, visible on each model's detail page where available.

Getting the most from it

Put stable content (system prompts, tool definitions, long context) at the start of the request and variable content at the end. Caches are prefix-based, so any change early in the prompt invalidates everything after it.