Architecting prompts for 10× cost reduction with caching

What gets cached

Anthropic caches the longest prefix of a prompt that ends at a cache_control breakpoint. The cache lives 5 minutes. Reads against a cache cost 10% of input rate; writes cost 25% above input rate. Breakeven is roughly two reads per write.

Where to put the breakpoint

System prompt, tool definitions, large documents, and example transcripts go above the breakpoint. The current user turn and any dynamic state — timestamps, session IDs, transient flags — go below.

The most common mistake is putting a timestamp or session ID in the system prompt for 'context'. Each call then invalidates the cache, and the operator pays full input rate for content that should have been 10%.

Multi-breakpoint structure

You can declare up to four breakpoints. The right structure for long-document workloads is: system → tools → static documents (breakpoint) → dynamic context (breakpoint) → user turn.

Cache misses in production

Watch the cache_read_input_tokens and cache_creation_input_tokens fields in the API response. A healthy long-running agent has cache_read 80% of input tokens. Below 50% means the cache architecture is fighting itself.