---
params:
  - omni?
---

:::note[How media inputs are billed]

Clef models convert media inputs to input tokens and bill them at the model's input token rate. Clef models do not charge for output tokens. For per-model rates, refer to [pricing](/workers-ai/platform/pricing/).

Images are tokenized as follows:

1. **Resize**: The image keeps its aspect ratio, but:
   - Each side is rounded to a multiple of 32 pixels.
   - If the area is under about 65,000 pixels (256x256), it is scaled up to reach that.
   - If the area is over about 1 megapixel, it is scaled down to fit.
2. **Count patches**: Each 32x32-pixel block is one token: `tokens = (width / 32) x (height / 32)`.
3. **Add markers**: 3 tokens are added to mark where the image starts and ends.

Each image is capped at 1,024 tokens.

{props.omni && (

<>
<p>Audio and video are tokenized by duration:</p>
<ul>
<li>
<strong>Audio</strong>: about 780 tokens per minute.
</li>
<li>
<strong>Video</strong>: up to about 15,400 tokens per minute at maximum
resolution, plus about 780 tokens per minute if the video has sound. A
480p clip uses about 8,600 frame tokens per minute (144 tokens per
second).
</li>
</ul>
<p>Each request accepts up to 4 audio clips and 2 videos.</p>
</>
)}

Media tokens count toward the model's context window, together with the questions. If they exceed the context window, the request fails. Otherwise, the text `state` is truncated to fit the remaining space.

:::