Prompt cache break-even calculator
Prompt caching charges more to write a prompt prefix to the cache and much less to read it back. A longer cache lifetime costs more to write, so it only pays when the prefix is reused often enough before it expires. Enter your prompt size and how your requests are spaced to compare the options.
Runs in your browser. Nothing you type or paste leaves this page.
Input cost for 30 requests
Cheapest: 5-minute cache, $0.228, 83% less than no caching.
| Option | Cost | Writes / reads | Pays off from |
|---|---|---|---|
| No caching | $1.32 | – | – |
| 5-minute cache | $0.228 | 1 / 29 | 2 requests |
| 1-hour cache | $0.258 | 1 / 29 | 3 requests |
Why
- 5-minute cache: a write costs 1.25× the input price and a read 0.05×. Each write adds 0.25×; each read saves 0.95×, so 1 read pays for a write. Each read keeps the entry for another 5 minutes, so a gap of 5 minutes or more means a new write.
- 1-hour cache: a write costs 2× the input price and a read 0.05×. Each write adds 1×; each read saves 0.95×, so 2 reads pay for a write. Each read keeps the entry for another hour, so a gap of an hour or more means a new write.
Prices from the providers’ pages, checked on 22 October 2026: Anthropic pricing and Anthropic caching docs, OpenAI pricing and OpenAI caching docs, Google pricing and Google caching docs. Standard rates in US dollars, before tax and any discounts. Check the current prices before you rely on an estimate.
How the estimate works
Each request sends the same prefix. The calculator walks through the requests in time order and, for each option, decides whether the prefix is written to the cache, read from it, or sent at the normal input price. The changing part of each request costs the same in every option. Output tokens aren’t included, because caching doesn’t change them.
Anthropic. A 5-minute write costs 1.25 times the input price and a 1-hour write twice it. Every cache hit refreshes the lifetime, so the entry lasts until 5 minutes or an hour after the latest request. Reads cost 0.1 times the input price on most models, 0.05 times on Claude Opus 5.5 and Claude Sonnet 5.5, and 0.025 times on Claude Fable 5.1.
OpenAI (GPT-5.6 and later, Responses API). The default 30-minute lifetime costs 1.25 times the input price to write, and a 12-hour lifetime 2.5 times. Reads cost 0.1 times the input price, or 0.05 times on GPT-6.1 Sol. A read extends 30-minute coverage to 30 minutes after that read, but doesn’t restart the 12-hour clock.
Google. Implicit caching has no write or storage charge, but a hit isn’t guaranteed, so it is shown as the best case. Explicit caching (not available on Gemini 4 Argon) stores the prefix for as long as you choose, billed per million tokens per hour. The calculator keeps it from the first request to the last, and counts creating it at the input price because the docs don’t give a separate price.
Break-even is the fewest requests, spaced as you entered them, for which an option costs less than no caching. For a single write, it takes the fewest reads whose combined discount is more than the write’s premium over the input price. Prefixes shorter than the model’s minimum aren’t cached at all. A request that arrives exactly when an entry expires is counted as a miss.