Image Token Cost Calculator
How many tokens an image costs a vision-language model under the common tiling rule — scale to fit, cut into tiles, a fixed base plus a fixed charge per tile — and what that means for a monthly volume.
Most multimodal APIs price images by tiles.
How the image token cost calculator works
Most multimodal APIs price images by tiles. The image is first scaled so it fits inside a maximum long side and short side, then divided into square tiles; the token count is a base charge plus a charge per tile. A 1024 × 1024 image at 512-pixel tiles is four tiles; a phone photo scaled to fit is usually six.
The practical lesson is that resizing before upload is the whole game: halving each dimension can cut the token cost by three quarters.
Formula: scale to fit (long ≤ L, short ≤ S); tiles = ⌈w/T⌉ × ⌈h/T⌉; tokens = base + tiles × per-tile
Worked examples
| Inputs | Tokens per image | Note |
|---|---|---|
| A 1080p screenshot | 1,105 | scaled to fit the short side |
| A square 1024 image | 765 | four tiles |
| A thumbnail | 255 | one tile |
FAQFrequently asked questions
How are image tokens counted?
Most vision APIs scale the image to fit a maximum size, cut it into fixed tiles, and charge a base amount plus a fixed number of tokens per tile. The exact numbers are inputs here because they differ by provider.
Why does resolution matter so much?
Because tiles are counted after scaling and rounded up. A 1024 × 1024 image is four tiles; a 1025 × 1025 image is nine. Resize before upload.
What is the base charge?
A fixed cost per image regardless of size, usually the cost of a low-resolution overview the model sees alongside the tiles.
Is there a low-detail mode?
Many providers offer one: the base tokens only, with no tiles. Use it when the model only needs the gist of the picture.
How does this compare with text?
A thousand tokens is about 750 words. A 1080p screenshot at full detail costs about as much as a page of text.
Where these figures come from
- Vaswani et al. (2017) — Attention Is All You Need — the transformer architecture the memory arithmetic follows
- Kaplan et al. (2020) — Scaling Laws for Neural Language Models — the compute relationship used for training estimates
- Hoffmann et al. (2022) — Training Compute-Optimal Large Language Models — the tokens-per-parameter guidance ("Chinchilla")
- IEEE 754 — Standard for Floating-Point Arithmetic — the numeric formats behind bytes per parameter
- NIST AI Risk Management Framework — the US federal AI framework
Last checked: September 2026. The relationships here are architectural, not vendor-specific: bytes per parameter follow the numeric format, KV-cache size follows the transformer definition, and token-per-word ratios come from published tokeniser behaviour.