Prompt Caching ROI Calculator

A cache hit is billed at 10% of the input rate on most models listed, 5% on Claude Opus 5.5 and 2.5% on Claude Fable 5.1. Writing to Anthropic's five-minute cache costs 1.25× input and its one-hour cache 2×, so a cached prefix pays off from the first read on the five-minute cache and the second on the one-hour cache. Google charges a storage fee per hour instead. Caching only works on an identical prefix, so put stable content first and anything that changes per request last.

Use the five-minute cache when requests reuse the prefix at least every few minutes, since each hit refreshes it and the write is cheaper. Use Anthropic's one-hour cache when the gaps are longer, for example a user who replies after twenty minutes or an agent step that runs past five: it costs 2× input to write and needs two reads to break even, but it survives the gap.

OpenAI lists a cache write rate of 1.25× input and a cached-input rate of a tenth, with no one-hour option, so the calculator treats it like a five-minute cache. Google's explicit context cache is billed at the normal input rate to create and then per million tokens per hour of storage, so it pays only above a steady request rate.

Caching matches an exact prefix, so any byte that changes invalidates everything after it. The usual culprits are a timestamp or request ID in the system prompt, tools listed in a different order, or JSON serialised with unsorted keys. Prefixes below a model's minimum cacheable length, from 512 to 4,096 tokens on Claude models depending on the model, are not cached at all. Check the cache-read token count each response reports.