Prompt caching savings calculator
See how much prompt caching saves on input tokens for 32 models, using each provider's official cached-input price.
Caching savings for every model
Same inputs, sorted by cost with caching. Input tokens only.
| Model | Input $/1M | Cached $/1M | Without caching | With caching | Saving |
|---|
Formula
cached = prompt tokens × sharewithout = requests × prompt tokens × input price ÷ 1Mwith = [requests × (prompt − cached) × input price + writes × cached × write price + (requests − writes) × cached × cached price] ÷ 1M
Write price is the model's cache-write price where our data has one, otherwise its normal input price. Prices last updated 2026-10-25, standard tier, short context. Notes on each model page list long-context, peak-hour and cache-write details.
Not sure your numbers are right? Get a $9 AI Agent Cost Audit
Most agent bills are dominated by context re-sent on every tool call, retries and output length, which a quick estimate misses. Send us your setup and get a written breakdown per task, day and month, cheaper model options and specific savings ideas within 48 hours.
Get my audit: $9 See a sample report
One-time $9 via Stripe Checkout. Nothing recurring. Prices from official provider pages.
FAQ
How is the saving calculated?
Without caching, every request pays the full input price for all prompt tokens. With caching, the cached prefix (prompt tokens x cached share) is billed at the cache-write price on cache writes and at the cached-input price on every other request; the rest of the prompt is always billed at the normal input price.
What is a cache write?
The first request with a new prefix (or the first after the cache expires) stores it in the provider's cache. If our data has no separate cache-write price for a model, writes are billed at the normal input price. Some providers charge more than the input price for cache writes, so check the model's source page.
Does this include output tokens?
No. Caching only changes the price of input tokens, so this tool compares input cost. Use the main calculator for full per-task cost.
How long does a cache last?
It depends on the provider, often 5 minutes to 1 hour, refreshed on each hit. If requests are spread out, the cache can expire and you'll pay for more writes; raise the cache writes number to model that.