What Is an AI Prompt Cache Hit Rate? (2026): The Share You Reused

An AI prompt cache hit rate is the share of input tokens, or of requests, that reused a cached prefix instead of paying the full price to read that prefix again. The feature that stores the prefix is what prompt caching is. The mechanism inside the model that keeps the already-read prefix is what a KV cache is. A product that bills a repeated prefix as its own object is what a context cache is. This page is the ratio: hits divided by the inputs you meant to cache. A high rate is a billing fact. It is not a quality score.

A cache that matches by meaning, which can return the wrong answer for a similar question, is what semantic caching for an LLM is. The unit on both the hit and the miss is what AI tokens are. Which team pays is what AI cost attribution is.

Prices below are US list prices. DeepSeek’s per-million figures are from the directory, checked October 16, 2026, for the deepseek-flash model. The directory does not record a per-token cache discount for Claude, ChatGPT, or Cursor seats, so this page does not invent one. Published October 24, 2026. We did not measure hit rates on a workload.

How to read the rate

Number	What it says	What it does not say
Hit tokens / input tokens you hoped to cache	How much of the repeated prefix was billed as a hit	That the answer was correct
Requests with any hit / all requests	How often the prefix matched at all	How much money you saved. A one-token hit is still a hit
Miss tokens	The prefix was rewritten, reordered, or too new to be cached	That the vendor is broken
The dollar gap	Hit price versus miss price, on a card that publishes both	A gap you estimated for a vendor that does not publish one

flowchart TD
  prefix["The same prefix is sent again"] --> match{"Does it match the cached prefix exactly?"}
  match -->|Yes| hit["Those tokens bill at the hit price"]
  match -->|No| miss["Those tokens bill at the miss price"]
  hit --> rate["Hit tokens divided by tokens you meant to cache"]
  miss --> rate
  rate --> read["A person reads the rate next to the invoice"]

What you put in the prefix, and in which order, decides whether the next call can hit. That ordering is what context packing is. A compacted prompt that changes every time will miss. That trade is what prompt compaction is.

A card that publishes both prices

DeepSeek’s directory record for deepseek-flash, which the card identifies as DeepSeek-V4.1-Flash, lists input prices per 1 million tokens. A cache hit is $0.003 off-peak and $0.006 at peak. A cache miss is $0.15 off-peak and $0.30 at peak. Output is $0.60 off-peak and $1.20 at peak. Peak hours on that card are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. The chat app is free. The full table is DeepSeek pricing explained.

On those figures, an off-peak miss costs 50 times an off-peak hit: $0.15 divided by $0.003. That ratio is arithmetic on the recorded card. It is not a promise that your traffic will hit. If the prefix changes every call, the rate is zero and you pay the miss price on every token.

Claude, ChatGPT, and Cursor sell seats, and the directory pages checked October 16 do not print a hit price and a miss price for those seats. Claude Pro is $20 a month, or $17 billed annually. ChatGPT Plus is $20 a month. Cursor Plus is $8 a month. Quoting a cache discount for those plans would be an invention. Read the API card you are actually billed on, and keep the seat price separate from the token price.

What to write down

Write the prefix you expect to hit, the rate formula you will use, the vendor prices you are using, and the date you read them. A hit rate without the miss price next to it is a percentage with no invoice. Review the rate when you edit the system prompt, because an edit is a miss until the new prefix is cached.

FAQ

Is a 90 percent hit rate a good model?

It means 90 percent of the tokens you counted reused a prefix, under the formula you wrote. The answers can still be wrong. Track quality on the eval rows you already have. Track the hit rate on the bill.

Do chat subscriptions have a hit rate?

Only if the vendor meters one. A Plus or Pro seat can include caching inside the product without showing you a hit price. Do not back into a rate from a seat price. If you need the rate, you need an API bill that separates hits and misses.

Does a semantic cache count as a hit?

Not on this page. A semantic cache returns a stored answer for a similar question. A prompt-cache hit reuses the computed prefix and still runs the new question. Similar questions are not the same request. Keep those rates in different columns.

What moves the rate?

A stable prefix, identical bytes, in the same order. Changing a date at the top of the system prompt, or inserting the user name before the cached document, can turn the rest of the prompt into a miss. Put the stable text first and the variable text after it.

Last verified

October 16, 2026 for DeepSeek, Claude, ChatGPT, and Cursor. Published October 24, 2026.

Sources: DeepSeek API pricing, Claude pricing, ChatGPT pricing, Cursor pricing.

Published by Andrew for AI Tools Optimizer.

DeepSeek 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.