RAG and data · 2026-09-22
Calculating cost for a mixed text+image collection on Voyage Multimodal 3.5
Covers Voyage Multimodal 3.5's two separate billing components (text per token, image per pixel) and the 50,000-to-2-million-pixel clamping behavior, for a collection containing screenshots and tables.
Billing has two separate components
Voyage Multimodal 3.5's catalog page explicitly splits billing into two components: text is priced per million tokens, and images are priced separately per billion pixels. That's a different cost model from pure text embedding models — estimating a collection's total cost requires knowing the ratio of text to image content.
On pages containing screenshots, tables, charts, and diagrams, the model's advantage is retrieving directly without converting the content to text first; but that advantage's cost depends on the pixel count of every image you send.
Account for the pixel-clamping behavior
The model page states that images below 50,000 pixels are upscaled to 50,000, and those above 2 million are downsampled to 2 million. That means even sending a very small icon or thumbnail bills at least a 50,000-pixel floor, and sending a very high-resolution image still caps the bill at 2 million pixels.
In practice, this means a collection sending many small icons can face a higher-than-expected floor cost, while a collection sending very high-resolution scans can see a lower-than-expected ceiling cost. Scaling images to these two thresholds before sending (not upscaling tiny ones unnecessarily, downsampling very large ones to a reasonable resolution) keeps total cost predictable.
- Below 50,000 pixels: billed at the floor, no cost advantage to sending an even smaller image.
- Above 2 million pixels: capped at the ceiling, no additional cost from sending an even higher resolution.
- Evaluate your collection's typical image resolution against these two thresholds.
Measure the text-to-image ratio, not an average
Assuming a single 'average document cost' when estimating a collection's total cost can be misleading; text-heavy pages and image-heavy pages carry very different cost profiles. Sampling your collection and measuring total text tokens and total image pixels separately is the most reliable way to estimate real cost.
Frequently asked questions
If I send text only, does image pricing kick in?
No, if your input has no image part, you're billed on text token price only; the per-pixel charge applies only to inputs that include an image.
Does sending a very small thumbnail lower the cost?
No, images below 50,000 pixels are still billed at 50,000 pixels; there's no cost advantage to going below that threshold.