Image Input Token Cost Calculator
Enter an image's width and height and see what it costs as input tokens under three different tiling formulas at once.
1,300 tokens per image under the selected formula. Across all three: 100–1,600 tokens, a 16.0× spread.
- A · size after both resizes
- 1,152 × 768 px
- A · tiles
- 3 × 2 = 6
- A · tokens (fitted tile grid)
- 1,300
- B · patches before the cap
- 2,035
- B · tokens (patch count)
- 1,600 — capped
- C · tokens (flat per image)
- 100
- Selected formula
- 1,300
- Cost per image
- $0.0039
- Roughly equivalent to
- 975 words of English text
- Monthly
- $780.00
- Annual
- $9,360
There is no such thing as the token cost of an image. There is the token cost of an image under a particular provider's tiling rule, and the rules differ enough that the same photograph can cost three times as much on one API as another at identical per-token rates. That is why this tool computes all three shapes at once and shows you the spread rather than presenting a single authoritative number.
The shapes are worth knowing on their own. A fitted tile grid charges a fixed base plus a rate per square tile after two resizes, so cost is roughly proportional to area up to the resize ceiling and then flat — which means uploading a 4000-pixel photograph costs exactly the same as resizing it to the ceiling yourself first, and sending far less data. A patch count charges per fixed-pixel patch up to a per-image cap, which makes cost proportional to area for small images and constant for large ones. A flat charge ignores size entirely, which is what a low-detail mode is: a promise that the model will only look at a thumbnail.
The practical consequences are the same under all three. Resize before you upload, because pixels above the provider's ceiling cost bandwidth and buy nothing. Crop to the region you actually need rather than sending the page and asking about a corner. And check whether a low-detail mode answers your question — for "is there a receipt in this photo" it usually does, at a tenth of the tokens. What this does not model: multi-page documents, where the page count multiplies everything, and video, which is billed per sampled frame.