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

# Billing model

> Prepaid wallet, pay-as-you-go consumption, metered to the micro-cent.

Omnia is **prepaid**: you top up a wallet, and every request is metered and
deducted from that balance in real time. There are no subscriptions, seats,
minimums, or idle GPU costs: you pay only for what you consume, and you can only
spend what you've funded.

## The wallet

Each workspace has exactly one wallet. You add funds with a card, and your
balance is debited as you use the platform. Balances are tracked internally in
**integer micro-cents** (1,000,000 micros equal \$1.00), so every charge is
exact and there is no floating-point drift, even across millions of tiny
per-request debits.

See [Managing your wallet](/billing/wallet) to top up and track your balance.

## What you're billed for

Omnia meters three kinds of consumption, each with its own pricing model.

<CardGroup cols={3}>
  <Card title="Per token" icon="message">
    Shared inference. Input and output tokens are priced separately, per
    million tokens.
  </Card>

  <Card title="Per GPU-hour" icon="microchip" href="/dedicated/overview">
    [Dedicated endpoints](/dedicated/overview), billed per-minute while the
    endpoint is running.
  </Card>

  <Card title="Per trained token" icon="graduation-cap" href="/fine-tuning/overview">
    [Fine-tuning](/fine-tuning/overview) jobs, charged once when the job
    completes.
  </Card>
</CardGroup>

<AccordionGroup>
  <Accordion title="Per-token (shared inference)">
    Requests to shared models are billed on the tokens they use. Input (prompt)
    and output (completion) tokens are priced independently, each quoted per
    million tokens. Before the request runs, Omnia places a hold and, when the
    request settles, replaces it with the exact charge. See
    [Managing your wallet](/billing/wallet) for how holds work.
  </Accordion>

  <Accordion title="Per-GPU-hour (dedicated endpoints)">
    A dedicated endpoint reserves GPU capacity for your workspace and is billed
    for the GPU-hours it runs, at **per-minute granularity**, only while it is
    running. Individual inference requests to a dedicated endpoint are **not**
    per-token metered; you've already paid for the GPU. Deploying requires
    enough prepaid balance to cover a minimum runway (default one hour), and an
    endpoint whose wallet runs dry is auto-stopped. See
    [Dedicated endpoints](/dedicated/overview).
  </Accordion>

  <Accordion title="Per-trained-token (fine-tuning)">
    A fine-tuning job is priced on the number of tokens processed during
    training and is charged **once, on completion**. See
    [Fine-tuning](/fine-tuning/overview).
  </Accordion>
</AccordionGroup>

## Pay-as-you-go, metered exactly

Usage is metered to the micro-cent and settled against your wallet as requests
complete. There are no estimates or month-end surprises; the amount you see
deducted is the amount the request cost.

For streaming requests, Omnia bills the tokens actually produced, **including
partial output when a stream is cancelled or disconnects**. You are charged for
what was generated, and never for nothing.

### Worked example

Suppose a shared model is priced at **$0.50 per million input tokens** and **$1.50 per million output tokens** (prices shown are illustrative). A request
sends a 1,200-token prompt and generates a 300-token completion:

| Component | Tokens | Rate (per 1M) | Cost                                       |
| --------- | ------ | ------------- | ------------------------------------------ |
| Input     | 1,200  | \$0.50        | 1,200 ÷ 1,000,000 × $0.50 = **$0.00060\*\* |
| Output    | 300    | \$1.50        | 300 ÷ 1,000,000 × $1.50 = **$0.00045\*\*   |
| **Total** |        |               | **\$0.00105**                              |

That total, 1,050 micro-cents, is what gets deducted from your wallet when the
request settles.

## Balance gating

Before a per-token request runs, Omnia checks your available balance and places
a short-lived **hold** sized to the request's worst-case cost: the prompt
estimate at the input rate plus the maximum possible output at the output rate,
priced at the sell price including markup. Maximum output is bounded by the
model's context window, so the hold can never be unrealistically large. If your
balance can't cover the hold, the request is refused with
`402 insufficient_balance` rather than letting you overspend. The hold is
released and replaced by the exact charge once the request settles.

<Note>
  A small **negative buffer** (down to at most −\$1.00) means a request already in
  flight won't fail mid-way over a rounding cent. But a wallet at or below that
  buffer stops accepting new requests. Keep a balance, or enable
  [auto-reload](/billing/auto-reload).
</Note>

## Auto-reload

Enable [auto-reload](/billing/auto-reload) to top up automatically from a saved
card when your balance drops below a threshold you set, so production traffic
never stalls on an empty wallet.

## Refunds

If a payment is refunded, the corresponding wallet credit is reversed
automatically.

## Seeing your spend

Every request writes a ledger entry with its tokens and cost, to the micro-cent.
Review it under **Usage** in the dashboard, or export it as CSV. See
[Usage & invoices](/billing/usage).