---
title: "What Is Prompt Caching and How Is It Billed?"
description: "A chatbot sends the same standing instructions, tool descriptions and earlier messages with every reply. Many model providers keep recently seen text in a short-lived cache and charge less to read it back. When the provider reports cached tokens for a reply, ProxyAI bills those tokens at that model's cache-read price instead of the normal input price, plus the usual ProxyAI margin. Some providers also charge a cache-write price the first time text is stored; where a model lists one, that is what is billed for those tokens."
category: "Quick FAQs"
canonical: "https://www.proxyai.app/help/quick-faqs/what-is-prompt-caching-and-how-is-it-billed"
---

> For the complete documentation index, see [llms.txt](https://www.proxyai.app/llms.txt). Request public content pages with `Accept: text/markdown`.

# What Is Prompt Caching and How Is It Billed?

A model with no listed cache price is billed at its normal input price. The Model Rates calculator assumes 80% of input is read from the cache from the second turn of a conversation, so its estimate is lower than a message-by-message guess with no caching. Actual savings depend on the provider and on how quickly a customer replies, because cached text expires after a few minutes.

### Read the provider-reported usage

Caching is an accounting detail of the model request, not a separate copy of your knowledge base that you manage in the dashboard. Repeated instructions may qualify for a provider's cache while new customer text still requires normal input processing. A long conversation can therefore contain both cached and uncached tokens.

Use the cached-token information in [Usage Log](/help/quick-faqs/how-do-i-review-or-export-usage) to understand a recorded event. Compare the selected model's input, output and cache prices on [Model Rates](/pricing). If a provider reports no cached input for a reply, do not assume the calculator's cache assumption applied to that event.

For example, two conversations with similar wording can produce different cache results when their timing differs or they use different models. Keep those circumstances alongside the token counts when comparing costs. The calculator is a planning tool; the log records what was reported and charged for the actual request.

For a wider explanation of billing units and comparisons, read [AI chatbot pricing models](/blog/ai-chatbot-pricing-models-explained).
