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

# Billing Model

> Queuing and environment setup are free; only the run stage is billed — unified per-second pricing for GPU and CPU

## How It Works

Every time you run a workflow or AI app on BizyAir, the task goes through three stages. **Only the "Running" stage hits your bill** — the other stages are free.

| Stage                       | Billable | Description                                                                                     |
| --------------------------- | -------- | ----------------------------------------------------------------------------------------------- |
| **Queuing**                 | Free     | The task waits in the shared queue for GPU scheduling — wait time is not billed                 |
| **Environment preparation** | Free     | Downloading model weights, pulling container images, initializing resources — all setup is free |
| **Running**                 | Billed   | Actual GPU / CPU time occupied by inference, settled per the billing model                      |

<Note>
  This means no matter how long the queue gets, how large the model is, or how thick the image is, only the part that is "actually running on the GPU / CPU" is billed.
</Note>

## Billing for AI Applications & ComfyUI Workflows

Upon submission, tasks enter a shared queue. The system automatically allocates instances based on the workflow's nodes and models used, meaning a single task's execution time may span multiple instance types.

Billable usage includes both GPU and CPU. All instances are billed per second. For concurrent tasks, charges accrue based on the cumulative execution time of each individual task, rather than the overall concurrent duration.

### GPU Billing Rates

| Instance | VRAM | Rate          | Specification Name      |
| -------- | ---- | ------------- | ----------------------- |
| RTX 4090 | 24GB | \$0.00050/sec | nvidia.rtx4090.g1.amd64 |
| RTX 5090 | 32GB | \$0.00085/sec | nvidia.rtx5090.g1.amd64 |
| H100     | 80GB | \$0.00250/sec | nvidia.h200.g1.amd64    |

### CPU Billing Rates

| Type | Rate           | Specification Name |
| ---- | -------------- | ------------------ |
| CPU  | \$0.000083/sec | cpu.c4m8.amd64     |

### Special Note

If a ComfyUI workflow uses Model API nodes (provided by the BizyAirCloudPlugin extension for ComfyUI), the execution time of these nodes will not be included in the CPU time billing. Billing for such nodes follows the Model API rules (see [Node Pricing](/pricing/node-price)).

### Mixed Workflow Billing

When a workflow combines BizyAir infrastructure nodes with Model API (third-party API / LLM / VLM) nodes, multiple fees accrue in parallel:

| Fee type                                                   | Billing method                                                 |
| ---------------------------------------------------------- | -------------------------------------------------------------- |
| Workflow infrastructure nodes (GPU + CPU)                  | Per-second by the allocated instance type                      |
| Model API nodes (e.g., Nano Banana, Kling, Sora, LLM, VLM) | Per-call or per-token, see [Node Pricing](/pricing/node-price) |

## Failure Rules

| Scenario                                  | Billable                           | Description                                                  |
| ----------------------------------------- | ---------------------------------- | ------------------------------------------------------------ |
| Node call succeeds, task ultimately fails | Billed                             | Any successful node call is included in billing              |
| User-initiated interruption               | Billed for the portion already run | Successfully run portions still incur charges                |
| Platform-caused forced interruption       | Free                               | Interruptions caused by the platform itself incur no charges |
| 5xx server errors                         | Free                               | Server errors are never billed                               |

## What's Never Billed

The following items will never appear on your bill:

* ✅ Queuing time — time a task spends waiting in the queue for GPU scheduling
* ✅ Environment setup time — downloading model weights, pulling container images, initializing setup
* ✅ Network transfer time — uploading input files and downloading output results
* ✅ 5xx server errors — any HTTP 500+ error is never billed
* ✅ Platform-caused interruptions — forced interruptions due to platform maintenance or failure

<Note>
  For long videos or multi-step sampling tasks, try a small-resolution dry run first before scaling up the output, to avoid wasting compute.
</Note>