What Is Prompt Caching? (2026): Reuse the Prefix You Already Sent

Prompt caching is a vendor feature that stores the computed form of a repeated prompt prefix, such as a system prompt or a long document, and charges less to reuse it on the next call. The prefix has to match. A reworded paragraph is a different prefix. Caching is a billing and latency lever. It is not a memory of your user’s last answer, and it is not a guarantee the model will say the same thing.

The unit on the bill is still what AI tokens are. The length of what you can attach is what a context window is. A cache that matches by meaning, which is a different mechanism and a different risk, is what semantic caching for an LLM is. The text you hope stays stable across versions is what prompt versioning is.

Prices below are US list prices from directory pages checked October 16, 2026. Published October 24, 2026. This page does not publish a per-token cache discount, because the directory does not record one for these seats. We did not benchmark latency.

What can hit, and what will miss

What you send	Cache behavior you can plan for	What breaks the reuse
A system prompt that is byte-for-byte the same	A prefix cache can reuse that block	You edited one sentence and shipped it
A long policy pasted at the top of every call	Same, if the vendor caches prefixes	The policy moved to the end, or a timestamp was prepended
The user’s new question	Usually the part you still pay to process	Treating the question as if it were free
Two questions that mean the same thing in different words	A prefix cache misses	That job belongs to a semantic cache, with a human rule for when a miss is safer

flowchart TD
  call["A request with a long stable prefix"] --> match{"Does the prefix match a stored block?"}
  match -->|Yes| reuse["The vendor reuses that block and bills the new suffix"]
  match -->|No| full["The call processes the prefix again"]
  reuse --> answer["You still read the answer"]
  full --> answer

A cache hit does not mean the answer was checked. It means the prefix was not recomputed from scratch. The person who accepts the answer is what human-in-the-loop AI is.

Where teams meet this, without a made-up cache price

A long Project. Claude Pro is $20 a month, or $17 billed annually ($200 up front). A Project that repeats the same instructions is the shape a prefix cache is built for, when the API or the product offers it. The seat price is not an API cache rate. The directory records Team as a per-seat plan and does not print a Team price. The review is Claude review. The container is Claude Projects vs ChatGPT Projects.

A repeating custom GPT or Project. ChatGPT Plus is $20 a month. Business is $20 a user a month billed annually, or $25 monthly, for two or more users, and does not train on business data by default. Put the stable instructions in the Project. Put the variable question after them. A date stamp at the top of the prompt will defeat a prefix match on purpose. The ladder is ChatGPT Go vs Plus vs Pro.

An editor that sends the same repo context. Cursor Plus for new individuals is $8 a month, or $80 a year, on The X Subscription. Repeated file context is a caching candidate only if the product documents a cache. Do not assume the seat includes one. The seat math is Cursor pricing explained.

The door for many API calls, when you have one, is what a model gateway is. The choice of which model runs is what AI routing is. Routing to a different model can change whether a cached prefix is valid. Read the vendor note before you budget a discount.

How to tell if it is working

Log the cache-hit field the vendor returns, if they return one. If they do not return one, you do not have a number, and you should not invent a savings percent for finance. Compare two calls that share a prefix against two calls that do not, on the usage export, and keep the raw rows. A chart without the raw field is a story.

Keep the prefix stable on purpose. Move dates, user names, and the question into the suffix. Version the prefix when you change policy, which is the point of prompt versioning. A cached old policy is a bug with a discount attached.

FAQ

Is prompt caching the same as chat memory?

No. Memory is a product feature that stores facts about a user. A prompt cache stores a computed prefix so the next identical prefix costs less to process. Turning memory on does not create a cache hit.

Will a cache make answers identical?

It can make the prefix cheaper. The new tokens are still generated. Temperature, a tool result, and the suffix can all change the answer. Do not use a cache as your regression test. That test is what AI eval testing is.

Should I cache user documents?

Only if your privacy review allows that vendor to retain the prefix for the cache window they document. A contract file in a cache is still a contract file. The seat’s training setting is a separate switch.

Does every plan include it?

The directory pages checked for this article do not list a cache add-on price. If the pricing card does not name caching, do not put a savings line in the budget. Confirm it on the API docs you will actually call.

Last verified

October 16, 2026 for Claude, ChatGPT, and Cursor prices named above. No cache discount rate is recorded in the directory for those seats. Published October 24, 2026.

Sources: Claude pricing, ChatGPT pricing, Cursor pricing.

Published by Andrew for AI Tools Optimizer.

Claude is in the chatbots and assistants directory. Directory pages may use affiliate links. A purchase through them may earn us a commission at no extra cost to you.