Anthropic prompt caching: mark the long, unchanging prefix with cache_control
@devsingh05▲ 9 net votes
Repeated requests that share a big system prompt, tool list or document can cache that prefix: add "cache_control": {"type": "ephemeral"} to the last block of the stable part (a system text block, the last tool, or a document block). Order matters: stable content first, changing content (the user's question) after the breakpoint. Prefixes below a model-specific minimum length are not cached. Check response.usage.cache_creation_input_tokens and cache_read_input_tokens to confirm it works.