Prompt caching (Claude API docs)

Summary

How to reuse a prompt prefix across requests so repeated context (system prompt, tools, documents, conversation history) is billed at a much lower cache-read rate and processed faster.

Key points

• Caching is a prefix match over tools, then system, then messages; any byte change invalidates everything after it.
• Mark cache breakpoints with cache_control, or use top-level automatic caching.
• Up to 4 breakpoints per request; prefixes below a model-specific minimum length are not cached.
• Check usage.cache_read_input_tokens to confirm hits.
• Common silent invalidators: timestamps in the system prompt, unsorted JSON, changing tool lists.

Why it matters

Usually the largest free cost and latency saving for agents and chat apps with long, stable context.