Prompt caching savings calculator
If your requests share a long, stable prefix — a big system prompt, a document, a tool schema — prompt caching can slash your input bill. Providers store that prefix and charge a deeply discounted rate when it is reused. This calculator models your cache-hit rate to show what caching actually saves on your workload.
How it works
Each request has two parts: the reusable prefix (the cached system prompt or context) and the new tokens unique to that request. Without caching, every request pays the full input rate on both parts.
With caching, a fraction of requests — your cache-hit rate — pay only the cheap cache-read rate (about 10 percent of input price) on the prefix, while the remaining requests pay a small cache-write premium to populate the cache. New tokens are always billed at the normal input rate. The calculator nets these out and reports daily and monthly savings.
Worked example
Imagine you run a legal-document assistant. Your system prompt — rules, document schema, and jurisdiction notes — is 4,000 tokens. Each request adds about 200 new tokens. You make 5,000 requests per day. At a 90% cache-hit rate:
- Without caching: 5,000 × 4,200 tokens × full input price each day.
- With caching: 90% of 5,000 requests pay the cheap cache-read rate on the 4,000-token prefix; 10% pay the higher cache-write premium to refresh it; all 5,000 pay the normal input rate on just the 200 new tokens.
The difference compounds quickly at this volume. The calculator works through exactly this arithmetic for your specific numbers and price inputs.
What affects your savings most
| Factor | Effect on savings |
|---|---|
| Prefix size | Larger prefix = more tokens cached = bigger saving per hit |
| Cache-hit rate | Direct multiplier — going from 60% to 90% hit rate roughly doubles savings |
| Daily volume | More calls means more opportunities to reuse the cached prefix |
| Provider rates | Cache-read discount depth varies — check current pricing |
Tips and notes
- Caching rewards big, stable prefixes. The longer and more reused the prefix, the larger the win. A tiny system prompt barely moves the needle.
- Watch the cache lifetime. Cached prefixes expire after a short window (minutes by default), so steady traffic keeps the cache warm; sparse traffic lets it lapse and lowers your effective hit rate.
- Combine with prompt compression. Shrinking the prefix before caching it saves on the write cost and slightly improves latency — the two techniques compound well.
- Rates are editable estimates. Cache-read and cache-write multipliers differ by provider and change over time — confirm current pricing before budgeting.