Prompt caching stores or reuses an internal representation of prompt content that a model has already processed. When later requests share the same prefix or context, the provider can avoid repeating part of the computation and charge a lower cache-read price.
Savings depend on workload structure, cache eligibility, retention, hit rate, and provider pricing. Long agent sessions with repeated instructions or documents can benefit substantially, while changing prompts or frequent cache misses may produce little saving.
ELI5
Prompt caching is a shortcut that lets an AI service reuse work it recently did on the same starting information. It can make repeated requests faster or cheaper when a large unchanged prompt is sent again.
For example, several coding requests may share the same long project instructions. A provider might cache that common beginning, but a new session or changed model may lose the shortcut, so the important instructions still need to be available in full.
What makes prompt caching useful?
It is most useful when many requests repeat a large, stable prefix such as system instructions, documents, examples, or an agent's working context.
Does prompt caching reduce every AI bill?
No. Savings depend on cache hits, eligible content, retention rules, workload repetition, and the provider's cache pricing.
