> ## 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.

# AI pricing

> What a model costs per million tokens, and who is charged.

AI is priced separately from minutes, on its own cards, per million tokens.

## The five dimensions

|                  |                                                    |
| ---------------- | -------------------------------------------------- |
| **Text input**   | Input that was text                                |
| **Audio input**  | The audio portion of input                         |
| **Cached input** | Input the provider had already seen — priced lower |
| **Output**       | What the model produced                            |
| **Audio output** | The spoken portion of the output                   |

<Warning>
  On a realtime speech-to-speech call, **audio output is usually the largest
  single line**. A price card that leaves it at zero understates every bill.
</Warning>

## Who pays

Set per credential:

* **`byo`** — the tenant's own key. The provider bills them directly. FireTone
  records the counts and no amount, because there is nothing for you to charge.
* **`rated`** — your key. Priced against the cards and added to their bill.

## Matching

A price is matched on provider and model, with an empty model acting as a
wildcard for that provider. An exact model beats the wildcard; a newer effective
date beats an older one.

**No match means no amount** — recorded as unpriced, never as zero. A zero would
assert the tokens were free.

## Approximate amounts

Where a provider does not itemise audio, the charge is flagged as approximate
and the audio is priced at the text rate. Google's realtime API is the case this
exists for: it reports two counts and no split at all.

## A virtual agent in a meeting

An agent sitting in a conference room as a default member holds one realtime
session for the whole meeting and is charged like any AI call: every minute
it listens is audio input, and every reply it produces is output -- including
replies nobody heard because the room had not addressed it (the seat records
how many, as `replies_unheard`). The minutes at the end are one text
completion by the agent's own provider on the organisation's key for it,
recorded as a charge of kind `minutes` against the session at the text
model's price on the same billing mode. It appears in the charges, not yet
on the AI report, which reads conversations.