Back to Tools

Prompt Caching Savings Calculator

A long system prompt that repeats on every call is a caching candidate. This calculator prices the cached portion at a discount and leaves the fresh user text at the normal input rate.

A long system prompt that repeats on every call is a caching candidate. This calculator prices the cached portion at a discount and leaves the fresh user text at the normal input rate.

Sample rates are planning examples, not official prices. Confirm the current price on the provider’s pricing page. This page is not affiliated with OpenAI, Anthropic, Google, or DeepSeek, and it is not financial advice.

How to use the Prompt Caching Savings Calculator

  1. Enter reusable system-prompt tokens and the tokens that change on every request.
  2. Enter output tokens, daily requests, and the share of requests that hit the cache.
  3. Set the cache discount. The default treats cached input as 90% off, which means you still pay 10%.
  4. Compare the uncached month with the cached month. Annual savings are that difference times 12.

How token math works

Providers bill input tokens and output tokens separately, usually per 1 million tokens. This suite estimates English prose at 1 token per 4 characters (about 0.75 words per token), code and JSON at 1 token per 3 characters, and multilingual text at 1 token per 2 characters. Dollar figures use the rate you select. They update in the browser and are not a quote from a model provider.

Practical notes

  • A 0% hit rate matches the uncached bill. A discount does nothing until requests actually reuse the cached prefix.
  • Output tokens cost the same in both columns. Savings come from the cached input only.
  • Next: AI Token Cost Calculator.
  • Also compare with the Batch API Discount Calculator.
  • Browse AI products in the AI tools directory when you want a tool listing rather than a token estimate.

FAQ

What is a cache hit rate?

The share of requests that reuse the stored system prompt instead of paying the full input price for those tokens. A 70% hit rate means 7 of 10 requests get the discount.

Is a 90% discount a real provider rate?

It is a default planning assumption. Providers publish their own cache price, minimum prompt length, and expiry. Replace the discount with the rate you were quoted.

Should I cache the user message too?

Only if that text repeats. Put one-off questions in the dynamic field so they stay at the full input price.