GPT Image 2.5 portrait 1024x1536 costs less than square

A 1024x1536 portrait image costs less than a 1024x1024 square on GPT Image 2.5, even though it has 50 percent more pixels. fal lists high quality at $0.04116 for 1024x1536 and $0.05268 for 1024x1024. Sume's x 1.25 makes that about $0.051 against $0.066 before input tokens. The cheapest listed size is 1024x768 at $0.03612.

The ladder at high quality

The output-token count in OpenAI's size and quality calculator is not proportional to pixel count: GPT Image 2.5 rounds the image to a grid of patches, and the grid for the long edge is shared across the ratio. The result is that near-square images cost the most per pixel. fal's page lists the prices below, and Sume's rate card reproduces them from the same formula.

Practical use

If the destination is a vertical or 16:9 placement, generate at that size instead of square and cropping: it is both the right shape and cheaper. Sume's image docs set the legal range: both edges multiples of 16, maximum edge 3840, aspect ratio at most 3:1 and between 655,360 and 8,294,400 pixels. Quality-by-size figures are in the full price table.