Input, output and cached tokens

By Seif Amara · Cost & pricing · Updated 2026-08-21

The three token types

Your AI bill is built from three separately-priced token types:

Input tokens: Everything you send: system prompt, conversation history, retrieved context and the user’s message. Usually the largest and cheapest per-token portion.
Output tokens: What the model generates. Priced higher than input, often 3–5×, so long responses cost more than long prompts.
Cached input tokens: Input the provider has seen recently and can reuse. Billed at a steep discount (often 10–25% of the normal input price). Worth it when a large, stable prefix (a long system prompt or fixed context) repeats across many requests.

When caching saves money

Set a cached-input percentage in the calculator to see the savings on your own workload. Not every model supports caching, and the calculator flags those. The bigger and more stable your repeated prefix, the more caching helps; a short or constantly-changing prompt gains little.

Key takeaways

Your bill is three separately-priced parts: input, output and cached input.
Output is usually priced 3–5× higher than input.
Cached input is billed at a steep discount when a stable prefix repeats.

FAQ

Does the system prompt count as input every request?

Yes. The system prompt and conversation history are resent and billed as input on every call, unless the provider caches a stable prefix at the discounted cached rate.

Why is output more expensive than input?

Generating tokens takes more compute than reading them, so providers price output several times higher than input.

Try it now

Put this into practice with the free calculator. No sign-up, runs in your browser.

Open the token calculator