3840x2160 YouTube thumbnail on GPT Image 2.5: valid size, 2 MB trap

Yes: 3840 x 2160 is a valid custom image size for GPT Image 2.5 on Sume. Both edges are multiples of 16 (3840 / 16 = 240 and 2160 / 16 = 135), the longest edge equals the 3840 maximum, the ratio is 16:9 (well under 3:1), and the pixel count is 3840 x 2160 = 8,294,400, which is exactly the documented upper limit of 8,294,400. YouTube recommends the same resolution for video thumbnails, so a native 4K thumbnail is possible in one call.

The part that bites is not the size. YouTube's help page says the file limit depends on where you upload: 2 MB for video thumbnails from mobile and 50 MB from desktop. A 4K PNG of a detailed scene can be much larger than 2 MB, so plan the export format before you generate a batch.

YouTube's thumbnail rules, as published

The values below come from YouTube Help (read 2026-10-09). The Sume side comes from the Image API docs.

Price of one 4K thumbnail

On Sume, GPT Image 2.5 at high quality in the 4K tier is $0.2225 per image (list price times 1.25, default options, output only). Ten thumbnails for ten videos cost 10 x $0.2225 = $2.225. Nano Banana 2.1 at its 4K tier is $0.20 per image, so ten cost $2.00. The catalog formula rounds billed amounts up to the cent, so treat both as planning figures and read usage.cost in the response.

If you do not need 4K, YouTube only states a minimum width of 640 pixels. Drafts at the 1K tier cost less: Nano Banana 2.1 is $0.10 at 1K and GPT Image 2.5 low is $0.02475, so you can test titles and faces cheaply and spend the 4K price only on the winner.

Keep the file under 2 MB

Sume documents png, jpeg and webp output for the models that accept the field, with webp the only option on Recraft V4. YouTube accepts JPG or PNG, so skip webp. JPEG is the practical choice when the 2 MB cap matters, because a photographic 4K frame compresses far better as JPEG than as PNG. Sume does not serve an output_compression field in v1, so you cannot tune the JPEG quality at generation time. Export or re-encode locally if the file is over the limit, or upload from desktop where the cap is 50 MB.

Large or slow requests may also come back as a 202 job envelope instead of the image. Check the status code, then poll the job result as described in the jobs docs.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "Wide thumbnail, surprised host left, bold empty space right for title",
  "quality": "high",
  "image_size": "3840x2160",
  "output_format": "jpeg"
}