YouTube thumbnail 1920x1080 is not a legal GPT Image 2.5 size

If you ask Sume for a 1920x1080 YouTube thumbnail from ChatGPT Image 2.5, the custom size is not legal: Sume's Image API docs require both edges of a custom image_size to be multiples of 16, and 1080 divided by 16 is 67.5. Three real 16:9 sizes do pass: 1280x720, 2560x1440 and 3840x2160.

This catches people because 1920x1080 is the size most creators reach for first, and it appears in the example sizes on fal's GPT Image 2.5 page. The rule you have to satisfy on Sume is the one in Sume's own docs, so check the size before you spend a request on it.

What YouTube asks for

YouTube's help page for thumbnails gives a single recommended resolution per thumbnail type and a floor below it. For regular videos that is 3840 x 2160 with a minimum width of 640 pixels, in a 16:9 ratio. You do not need to hit the top size: anything 16:9 above the floor is within the page's numbers, and the file-size limits are what usually bite on a phone upload.

The same page sets 2 MB as the limit for video thumbnails uploaded on mobile and 50 MB on desktop. A 4K PNG of a busy thumbnail can be large, so the smaller sizes below are often the practical choice.

The size rule on Sume

The docs list four limits for a custom GPT Image 2.5 size: both edges multiples of 16, a maximum edge of 3840, an aspect ratio of at most 3:1, and a total of 655,360 to 8,294,400 pixels. Run the 16:9 candidates through them and only the sizes whose short edge divides by 16 survive.

If you cannot use a custom size, the docs also list named presets and a 16:9 entry for aspect_ratio. The 1920x1088 row below is not exactly 16:9 (it is 1.7647 against 1.7778), so crop 8 pixels afterwards if you need the exact ratio.

Check the size, then send it

This script applies the documented limits to your candidate sizes and then sends the 2560x1440 request. It prints the HTTP status because the docs say to check the status code, not the body shape: 200 is the image response and 202 is the job envelope you then poll.

import os
import requests


def legal(w, h):
    px = w * h
    return (
        w % 16 == 0 and h % 16 == 0
        and max(w, h) <= 3840
        and max(w, h) / min(w, h) <= 3
        and 655_360 <= px <= 8_294_400
    )


for w, h in [(1280, 720), (1920, 1080), (1920, 1088), (2560, 1440), (3840, 2160)]:
    print(w, h, w * h, legal(w, h))

resp = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"]},
    json={
        "model": "openai/gpt-image-2.5",
        "prompt": "Bold YouTube thumbnail, surprised host beside a glowing laptop",
        "image_size": {"width": 2560, "height": 1440},
        "quality": "high",
    },
    timeout=60,
)
print(resp.status_code)  # 200 = image body, 202 = job envelope

Quality and cost

Quality and size both move the price. The fal page lists a range from $0.00402 for 1024x768 at low quality to $0.40026 for 3840x2160 at max, and says larger images, longer prompts and more complex requests cost more. Those are fal's list figures, not what Sume charges. Sume's docs say omitted quality defaults to high, that auto quality reserves max, and that auto size reserves the output-token upper bound, so pass an explicit quality and an explicit legal size when you want a predictable hold.

For a thumbnail that will be shown at a few hundred pixels wide, 1280x720 or 2560x1440 at high is usually enough to review. Keep 3840x2160 for the winner, and use the same prompt so the final matches the draft you picked. Slow configurations such as 4K and high quality are the ones most likely to come back as a 202, so be ready to poll the job as described in Jobs and results.