Prompt Caching Guide

What prompt caching is

Prompt caching lets a provider store a prompt prefix it has seen recently and bill it at a discounted cached-input rate on later calls. Instead of paying the full input price for a long system prompt on every request, you pay full price once and the discounted rate after that, for as long as the prefix stays unchanged.

When it pays off

Caching helps when two conditions hold: the prefix is large, and it repeats across many requests. A long, fixed system prompt or a stable block of retrieved context that every request shares is the classic win. A short prompt, or one that changes on every call, gains little because there is little to reuse.

How much you can save

Savings depend on what fraction of your input is cacheable and how big the cached discount is. Estimate the cacheable share (usually the system prompt) and the prompt caching calculator applies the cached rate to that fraction for you. Background on the token types is in the input, output and cached tokens guide.

FAQ

What is cached input?

Input tokens the provider has seen recently and can reuse, billed at a lower rate than normal input. It applies where the model supports prompt caching.

How much cheaper is cached input?

Discounts vary by provider and are often a fraction of the normal input price. Check the cached rate on the pricing table for each model.

Does caching change the model’s output?

No. Caching only changes billing for a reused prefix. The response is generated the same way.