Prompt Caching Calculator

Cache reads cost 2% to 10% of the standard input price on current models, so the repeated span of a prompt costs 90% to 98% less once it is cached. What you actually save depends on how often a request finds the cache warm. The calculator works that out from your request rate and the cache lifetime, then charges the read price on hits and the write price on misses to give a monthly figure for your own numbers.

Providers that charge a write premium (Anthropic 1.25x on the 5-minute cache and 2x on the 1-hour one, GPT-5.6 1.25x) only save money once enough requests find the cache warm. With a 1.25x write and a 0.1x read, break-even is a 22% hit rate, which a 5-minute cache reaches at about 3 requests an hour. The 1-hour cache needs a 53% hit rate but gets there at under 1 request an hour. Providers with no write fee (DeepSeek, Gemini implicit caching) save from the first hit.

The stable content at the start of every request: your system prompt, tool and function definitions, few-shot examples, and any fixed reference documents. It has to be identical byte-for-byte across requests to hit the cache. The variable part, the user message, always sits at the end and is never cached, which is why it does not affect the saving.

Yes, for cost. Every hit is billed at the read discount instead of full price, and on Anthropic and OpenAI each hit also resets the cache timer for free, which is why busy traffic keeps a cache warm almost indefinitely. The usual reason hit rates are low is a dynamic element (a timestamp, session ID or rotating instruction) buried in the prefix, which changes the hash and forces a fresh write every request. Move anything variable to the end of the prompt.

Every model in the site price list: Claude (Fable, Opus, Sonnet, Haiku, with 5-minute and 1-hour caches), GPT-5.6 (Sol, Terra, Luna, 30-minute cache), Gemini (implicit caching, and explicit caching with its storage fee) and DeepSeek, each with its own read price, write fee and minimum cacheable prompt, checked against the providers in September 2026. Pick "Custom" to enter the terms from your own contract.