Multimodal Image Tokens Calculator
Inputs
| Width | 1,024 |
|---|---|
| Height | 1,024 |
| Detail | High detail |
| Base Tokens | 85 |
| Tokens per Tile | 170 |
Multimodal Image Tokens Calculator
Estimate how many tokens a vision model charges for an image from its pixel dimensions and detail setting, using the documented resize-and-tile accounting.
Inputs
Image
Token constants
Results
Enter a value to see results.
Details
Multimodal image tokens
When a vision-capable language model reads an image, it does not charge a flat fee — it converts the image into a token count and bills those tokens at the same rate as text. The count depends on the image's pixel dimensions and on the requested detail level, following a documented resize-and-tile procedure. This calculator reproduces that procedure so you can estimate the token cost of an image before sending it, which matters because a single high-resolution image can cost more tokens than a paragraph of text.
How the count is built
In high-detail mode the model resizes the image in two steps and then covers it with square tiles. First it scales the image down to fit inside a 2048-pixel square if either side exceeds that, preserving the aspect ratio:
s1=min(1, max(w,h)2048)Then it scales again so the shorter side becomes 768 pixels:
s2=min(w⋅s1, h⋅s1)768The resized image, of dimensions , is divided into a grid of 512-pixel tiles, rounding up on each axis so partial tiles still count:
k=⌈512w2⌉⋅⌈512h2⌉Each tile costs a fixed number of tokens, and a flat base cost is added to cover the low-resolution overview the model always receives:
timg=t0+ttile⋅kIn low-detail mode the tiling is skipped entirely and the image is billed at the base cost alone, .
Worked example
Take a 1024×1024 image in high detail, with a base cost of 85 tokens and 170 tokens per tile. Neither side exceeds 2048, so the first scale leaves it unchanged; the second scale brings the shorter side to 768:
s2w2=h2=1024768=0.75=1024×0.75=768 pxThe tile grid is
k=⌈512768⌉⋅⌈512768⌉=2×2=4and the total is
timg=85+170×4=765 tokensAt a mid-tier input rate of three dollars per million tokens, that single image costs about 0.0023 dollars — modest alone, but a batch of thousands of images quickly dominates a text-only budget.
Notes and variations
Because the shorter side is fixed at 768 pixels before tiling, any image already at or above that size produces the same grid: a 4096×4096 photo bills the same 765 tokens as the 1024×1024 example above. Sending higher resolution buys no extra detail past the cap and no extra cost. The exact constants — base cost, tile size, the 2048 and 768 thresholds — differ across providers and model versions, so treat the defaults here as one documented scheme rather than a universal rule.
Application
Image tokens feed straight into per-call pricing: add them to the prompt's text tokens and price the total with the Token Cost Calculator. When the images themselves are generated rather than read, their cost is estimated separately by the Image Generation Cost Calculator.
Frequently Asked Questions (FAQ)
How does image tiling determine the token count?
The model first resizes the image: it scales the image down to fit within a 2,048-pixel square if either side is larger, then scales again so the shorter side is 768 pixels. The resized image is divided into a grid of 512-pixel tiles, and each tile costs a fixed number of tokens.
A flat base cost is added on top to cover the low-resolution overview the model also receives. The total is the base cost plus the per-tile cost times the number of tiles.
What is the difference between low and high detail?
High detail runs the full resize-and-tile procedure, so the cost grows with the image size and the model can read fine elements such as small text. Low detail skips tiling entirely and bills only the flat base cost, giving the model a single coarse overview at a fixed, lower price. Low detail suits images where only the overall content matters, such as a chart shape or a scene category.
Does sending a larger image cost more?
Up to a point. Because the image is resized so its shorter side becomes 768 pixels before tiling, any image already at or above that size produces the same tile grid and the same token count. Sending a 4K photo instead of a 1,024-pixel one therefore costs no more in high detail, while a small image below 768 pixels on its shorter side may produce fewer tiles. The exact cap and tile size depend on the model.
Disclaimer
Vision token accounting varies by provider and model, and providers revise it over time. The defaults follow one documented scheme; confirm the current base cost, tile size, and resize rules for your specific model before relying on a figure for billing.