Prompt caching stores the processed state of a prompt prefix on the provider's side. When a later request shares that prefix exactly, the provider can skip the recomputation and charge a lower rate for the reused tokens. It is most useful when a long system prompt, tool schema, or document preamble repeats across many requests.
The pricing is asymmetric. At the providers we checked, a cache read costs about a tenth of the base input price, while a cache write costs about 1.25 times the base for a short-lived cache and about twice the base for a longer one. A single reuse usually pays for a short-lived write; a cache that expires between requests does not.
Caching rewards stable order. Put the fixed material first and the volatile material last, and do not inject timestamps or shuffled search results into the prefix. Measure the hit rate; a cache you assume is working is a line item you cannot explain. See estimate the token bill.