Module cache 

Cache breakpoints for Anthropic prompt caching.

Where to place breakpoints

A cache_control: ephemeral marker tells the server “cache the request up to and including this block.” Typical strategies:

• Cache system + tools + first user turn. The longest-lived prefix. Use CreateMessageRequestBuilder::cache_control_on_system, CreateMessageRequestBuilder::cache_control_on_tools, and either CreateMessageRequestBuilder::cache_control_on_last_user (called once before the first send) or ContentBlock::text_cached to mark that user turn at construction time.
• Refresh on each turn. For long conversations, place a fresh ephemeral breakpoint on the most recent user turn each request. Conversation::with_auto_cache(AutoCacheMode::SystemAndLastUser) does this for you.

TTL choice: the default 5-minute cache is right for nearly all interactive workloads. "1h" requires the extended-cache-ttl-2025-04-11 beta header and is meant for batch / long-running pipelines where the same prefix sees sustained traffic.

Enums

CacheControl

Marks a cache breakpoint on a content block, system prompt, or tool definition.