OpenAI Prompt Caching Calculator
Find out when reusing a prompt prefix saves money.
Compare ordinary input with one warm-up write and later cache reuse.
Last verified: 2026-09-24
The first request writes the prefix; later hits read it, while misses write it again. GPT-6 needs a 1,024-token eligible prefix. Its minimum 30-minute lifetime restarts after each write or reuse; routing can still cause misses.
What the caching estimate includes
Enter the stable prefix shared across your requests, followed by the fresh input and expected output for each request. The first request writes the prefix. For later requests, the selected hit rate divides traffic into reads and new writes. Fresh input and output are charged on every request.
A GPT-6 cache write costs 1.25 times ordinary input; a cache read costs 10% of ordinary input. The write premium makes a one-off prompt more expensive, while a successful reuse can repay that premium. The calculator compares both routes across the same number of requests and applies the selected model’s service mode and long-context rules.
Worked example: ten requests to Astra
With a 10,000-token reusable prefix, 1,000 fresh input tokens and 2,000 output tokens per request, ten ordinary Standard requests cost $2.10. One prefix write followed by nine successful reads costs $1.315, saving $0.785. Fresh input and output account for the unchanged part of the bill.
If those later requests miss the cache, they write the prefix again. At a zero hit rate, this workload costs $2.35. The break-even rate after the initial write is about 24.15%. Fractional hit counts represent an expected workload average; actual requests have a read or a miss.
Cache lifetime and reuse
GPT-6 requires at least 1,024 eligible visible input tokens in a cacheable prefix. The default prompt_cache_options.ttl is 30m. A write or reuse starts another minimum 30-minute eligibility period, and a read refreshes it without another write charge. Expiration and routing affect whether the next request hits; the calculator takes your expected hit rate rather than promising one.
Place stable instructions, tool definitions and shared reference material before changing content. Keep the prefix unchanged between requests. Editing early text can prevent later content from matching, even when much of the prompt looks similar to a reader.
Read the break-even result
The tool includes the initial write when finding the hit rate needed to recover the premium. With a large request count, the threshold approaches 21.74% for the verified GPT-6 rates. With fewer requests, the warm-up write has a larger effect. One successful reuse pays for one initial write premium when both requests use the same tier and prefix.
The estimate assumes that eligible misses write the full reusable prefix and that fresh input is billed as ordinary input. Use actual API usage fields for a completed run: input totals, cached tokens and cache-write tokens should be separate categories of the same input, not added twice.
Read the complete GPT-6 pricing guide or calculate a single request when you already have those usage fields.