GPT Image 2.5 at 3840x2160: low $0.014 to max $0.50 per image

A 3840x2160 image from ChatGPT Image 2.5 costs $0.0140 at low and $0.5004 at max on Sume, with medium at $0.0325 and high at $0.1251. The size is valid as a custom image_size because 3840 and 2160 are both multiples of 16 and the pixel count, 8,294,400, is exactly the allowed maximum.

The figures are the repository estimate from the Fal token rates the docs cite, and Sume's catalog row is the number that bills.

The ladder

One row per quality for a single text-to-image call, no references.

Why quality matters more than size

At this size the jump from low to max is about 36 times. Resolution alone does not set the bill; the quality tier does. At 1024x1024 high is $0.0659, so a 4K low image at $0.0140 is cheaper than a square high one.

If you leave quality as auto, Sume reserves the max price for the call, and the default when you omit the field is high. Neither is what most 4K jobs need.

• Use low or medium for layout tests.
• Use high for finals you will crop and print.
• Reserve xhigh and max for text-heavy art you cannot redo.

When 4K is worth paying for

Few images need 3840x2160 pixels. A 4K master helps when you will crop into it for several formats, print it large, or show it on a screen where fine detail is visible. For a web page that displays it at 1200 pixels wide, a 1536x864 high image at $0.0405 looks the same to the reader and costs a third as much as the 4K high image at $0.1251.

A better pattern for most teams is to draft at low in 4K only when composition at that size matters, and otherwise draft small. Approve the framing at the cheap size and then regenerate the single approved prompt at 4K high. One final 4K high image plus five low drafts costs about $0.20 at 3840x2160.

• Prefer a smaller size for drafts.
• Use xhigh or max only when text accuracy at 4K matters.
• Expect a 202 and build the polling path first.

Expect a 202

The docs name 4K and high quality as the settings most likely to exceed the 30-second wait. When that happens, POST /v1/images returns 202 with a job envelope, and you poll GET /v1/jobs/{id}/status and then GET /v1/jobs/{id}/result. Check the status code before you read the body. The Fal page lists the token rates.

This ladder describes one text-to-image call with no references. Input image tokens add to the total at $8 per million and are estimates, so confirm the real figure from usage.cost after the first call.

If you need many 4K images, spread them across calls rather than raising n, because the docs list large n among the settings that most often push a call past the 30-second wait.