Prompt Caching Savings Calculator

Why the cached prefix, not total tokens, is what decides the savings

Prompt caching only discounts the part of a request that's identical to a recent previous call — usually a system prompt, a tool schema, or a big shared context block pasted at the top of every request. The unique part (the actual user message, the part that changes every time) never benefits, no matter how the cache is configured. That's why the two numbers that matter most here are the split between cached and unique tokens, and the cache hit rate: a large shared prefix reused across thousands of near-identical calls (a fixed system prompt, a long document being queried repeatedly) can cut input cost by most of what caching discounts, while a workload with little repeated context barely moves.

Hit rate matters because caches expire — most providers keep a cache warm for a few minutes, so a burst of requests using the same prefix hits it repeatedly, while sparse, spread-out traffic keeps paying full price to refill it. Providers also differ in mechanics: some cache automatically and discount by default, others require you to mark the cached portion explicitly and may charge a small premium the first time they write it. Check your provider's current caching docs for the exact numbers; the preset here gets you in the right ballpark for planning.

FAQ