> ## Documentation Index
> Fetch the complete documentation index at: https://docs.firetone.com.au/llms.txt
> Use this file to discover all available pages before exploring further.

# What AI costs

> Tokens, where they are counted, and how to read a figure that is approximate.

Every AI call records what it used. You see it on the call record, on the
conversations screen, and in the billing summary.

<img src="https://mintlify.s3.us-west-1.amazonaws.com/firetone/images/call-ai-panel.png" alt="What the AI did, on a call record" />

## The counts

|              |                                |
| ------------ | ------------------------------ |
| **Text in**  | Input that was text            |
| **Audio in** | The audio portion of the input |
| **Out**      | What the model produced        |

<Warning>
  **Audio in is part of the input, not a third total.** Adding the three
  together counts the audio twice. FireTone does the subtraction in one place so
  a report, a settlement and a screen cannot disagree — but if you are adding
  these up yourself, do not.
</Warning>

## When a figure is approximate

Some providers do not itemise audio at all. Where that is so, FireTone says the
amount is approximate rather than presenting it as exact.

<Note>
  Google's realtime API reports a prompt count and a response count and nothing
  else. Its audio is therefore priced at the **text** rate, because there is no
  other number to price it at. "No audio reported" and "no audio" are the same
  zero, and the product refuses to let you read one as the other.
</Note>

## Who pays

Two arrangements:

* **Your own key.** You are billed by the provider directly. FireTone still
  records the tokens, because "what did this profile consume" is worth
  answering whoever paid.
* **The operator's key.** Priced against a rate card and added to your bill.

An amount is **absent, never zero**, in three situations that must not be
confused: you may not see money, no price matched the model, or your own key
paid. Each says which.

## Prices are frozen at the call

A call is priced against the rate that applied when it ran, and the numbers are
snapshotted onto the charge. A price changed today does not restate what last
month cost.