Workspace & BillingAI tools

How AI calls are billed

An AI call costs a base charge plus the tokens it uses. Here is exactly how that is calculated, what a token is, and how to keep the cost down.

Updated August 9, 2026

Two parts: a base, then tokens

Every AI call is charged as:

base credits  +  credits for the tokens used

The base is fixed per tool and reserved before the call runs. The token charge is added afterwards, because nobody — including us — knows how many tokens a model will use until it has finished.

What a token is

A token is roughly four characters of English. A 1,000-word document is around 1,300 tokens. Both directions are counted:

  • Input tokens — everything sent to the model: your text, our instructions, and any instructions you have saved.
  • Output tokens — what the model writes back.

Output is the expensive half. Every provider charges several times more for it, so the single biggest lever on what a call costs you is how long an answer you ask for.

What the rates are

Every model sits on a tier, and the tier sets the rate:

TierCredits per 100 input tokensPer 100 output
Economy37
Balanced513
Advanced1024
Premium2570
Embedding3

The tier is shown wherever you choose a model, so you always know what you are agreeing to before a call runs. Output costs more than input on every tier because it costs more to produce — that is why asking for a shorter answer is the most effective way to spend less.

Embedding models charge for input only. There is no generated output, so the cost of an embedding call is entirely predictable from what you send.

The ceiling

Every AI tool has a maximum credits per call. If a call somehow exceeds it, you are charged the maximum and not a credit more — we absorb the difference. You will never receive a surprise bill from a single request.

Keeping the cost down

  1. Ask for shorter output. The summary-length control on the web tool, or maxLength over the API.
  2. Send less. Trim boilerplate before summarising; it is charged as input.
  3. Keep saved instructions short. Anything saved in your API configuration is sent with every call, so a 500-word instruction is a permanent per-request cost.
  4. Use a cheaper model where quality allows. See Choosing a model.

Failed calls

If a call fails, you are not charged. The reservation is returned, and the attempt is recorded in your usage with a cost of zero — so you can see that it happened and confirm you were not billed for it.

Related