Anthropic prompt caching reduces repetition in long AI contexts

The cache reuses stable prompt prefixes to reduce latency and cost. Understand TTL, breakpoints, invalidation and metrics for production.

Caching requires separating the stable from the variable

Policies, examples and tool definitions must appear before data that changes with each request. A timestamp inserted into the prefix can prevent all hits.

TTL keeps pace with the flow

Five minutes attend active conversations; an hour can serve spaced analyses, with a higher writing cost. The choice needs to consider real cadence and repetition.

Metrics confirm the gain

Cache read and cache creation must be monitored along with latency and cost. Without cache reading, the optimization exists only in the code, not in the result.

Tools also change the prefix

Changing settings or their order invalidates subsequent layers. Dynamic catalogs must use mechanisms that preserve the stable set when the platform supports it.

Nexus Reading

Prompt caching is context engineering: stabilizing contracts, measuring hits, and maintaining privacy. The sustainable economy depends on predictable architecture, not on activating an isolated field.

FAQ

How long is the cache?

The default is five minutes and there is an option for one hour, according to the current documentation.

How do I know if there was a cache hit?

Check cache_read_input_tokens; cache_creation_input_tokens shows content written to the cache.

Does cache store prompt text?

Anthropic reports that KV representations and hashes are in memory during TTL, without storing the raw text at rest.

Essential guides to delve deeper into the decision