Upscale a 1280x720 thumbnail to 3840x2160: factor 3 or generate at 4K?
Factor 3 turns 1280x720 into 3840x2160 for $0.20; GPT Image 2.5 at 4K is $0.2225. For 36 thumbnails: $7.20 to upscale, $8.01 to regenerate.
Sume image upscale with upscale_factor: 3 turns a 1280 x 720 thumbnail into 3840 x 2160 for a flat $0.20. Generating a new image at GPT Image 2.5 4K costs $0.2225, so for 36 finished thumbnails the upscale is $7.20 and a regeneration is $8.01. If the thumbnails already exist, upscale. If they do not, generate at the larger size directly.
What YouTube asks for
YouTube's help page for custom thumbnails recommends 3840 x 2160 pixels for videos and 2160 x 3840 for Shorts, in JPG or PNG. It also says to stay under 2 MB when you upload from a mobile device, and 50 MB from a desktop. The page adds that the image should be as large as possible because it is the preview in the embedded player.
A 3840 x 2160 PNG with flat graphics and text is often well under 50 MB, but it can exceed 2 MB. Ask for output_format: "jpg" when the file has photographic content and the upload goes through a phone.
| Route | Per thumbnail | 36 thumbnails |
|---|---|---|
| Upscale existing 1280 x 720 files, factor 3 | $0.20 | $7.20 |
| GPT Image 2.5 at the 4K row, new image | $0.2225 | $8.01 |
| Low draft ($0.02475) then upscale | $0.22475 | $8.091 |
| Medium draft ($0.055625) then upscale | $0.255625 | $9.2025 |
Why a draft plus upscale is not cheaper
It is easy to assume that a cheap draft followed by an upscale undercuts a 4K render. The table shows it does not. A low draft plus the $0.20 upscale already costs $0.22475, which is above the $0.2225 direct row. The upscale only pays when you cannot or do not want to regenerate: the thumbnail is a photo, or a design that you have approved, or text that must not change.
Regeneration also changes the picture. Even with the same prompt, the model draws a new image, and a thumbnail that tested well is not the same thumbnail after a re-render.
Request
The upscale route takes image_url, upscale_factor from 1 to 4 and an output_format of png, jpg or webp. The catalog lists the price as $0.20 per image. Sume's note on the price reserves about 16 megapixels of output, and a 3840 x 2160 result is about 8.3 megapixels, well inside that. See the reserve explainer for what the hold means.
Keep the 1280 x 720 originals. If a video later needs a vertical Shorts version, you will be starting from the source again, and the 2 MB vs 50 MB post covers the file-size side.
{"image_url": "https://example.com/thumbs/ep-114.png",
"upscale_factor": 3,
"output_format": "jpg",
"mode": "async"}Checks before you upload 36 files
Open three results at full size and look at text edges, because the headline is what people read first. Confirm the dimensions are 3840 x 2160, the format is what you asked for, and the size on disk is under 2 MB if you will upload from a phone. Pillow can do all three in a few lines.
Keep a record of the job id per video and the factor you used. If YouTube's guidance changes or you want a vertical version, you can start again from the original at a known cost of $0.20 each.
Sources
Related posts
More in Pricing
- Voice API cost gap up to 213x: what drives it, how to price yours
Krisp's newsletter reports a voice API cost gap of up to 213x across OpenAI, Gemini and Qwen. Why list TTS prices differ far less, and how to normalise yours.
- Wallet reserve for a full queue of Seedance 2.5 clips, by plan
Sume reserves the estimate at submit. A full queue of 5 s Seedance 2.5 720p clips ($2.889 each) reserves $17.33 on Free and $346.68 on Scale.
- Wan 3.0 on Sume: a 2-second minimum clip and a 30-second max, priced
Wan 3.0 accepts 2 to 30 seconds. On Sume a 2 s clip is $0.125 at 480p and $0.50 at 1080p; 30 s at 1080p is $7.50.
- Wan 3.0 for 500 prompts: 480p, 720p or 1080p costs 1x, 2x or 4x
On Sume, a 500-prompt Wan 3.0 library costs $250.00 at 480p, $500.00 at 720p and $1,000.00 at 1080p for 8-second clips. Resolution sets the bill.
Written by Sume