Prompt Caching Savings Calculator (cache writes, hits & break-even reuse)

How it works

Without caching every request pays the full input price for the prefix: requests × prefix tokens × price. With caching, the requests that write the cache pay the write price (input price × write multiplier) and all other requests pay the hit price (input price × hit multiplier). Each write costs (write multiplier − 1) extra and each hit saves (1 − hit multiplier), so caching pays off once hits per write exceed (write − 1) ÷ (1 − hit). With Anthropic’s 5-minute cache (1.25× write, 0.1× hit) one hit per write is already enough. Only the prefix is counted here; the rest of each prompt and the output cost the same either way.

⚠ Estimates for information only — not financial advice.

FAQ

How do I know how many cache writes I have?

API responses report cache-creation and cache-read tokens separately. Divide a day’s cache-creation tokens by the prefix size.

Why might my cache not be hit?

Caches match the exact beginning of the prompt. Put fixed content (instructions, tools, documents) first and changing content (the user’s question, timestamps) last, and keep requests within the cache lifetime.

Checked 10 Nov 2026. Cache write and hit multipliers as published on the Anthropic pricing page (1.25× / 2× write, 0.1× hit) and the cached-input prices of OpenAI, Google, DeepSeek and xAI, checked 10 Nov 2026.