15 thumbnail tests (5 videos x 3 variants): $2.25 at 2K on Sume

The short answer

Fifteen thumbnail images (5 videos x 3 variants) on Nano Banana 2.1 through Sume cost $1.50 at 1K, $2.25 at 2K and $3.00 at 4K. The Sume prices are $0.10, $0.15 and $0.20 per image, and the arithmetic is 15 x 0.10, 15 x 0.15 and 15 x 0.20. Ask for aspect_ratio: "16:9", which YouTube recommends for video thumbnails.

What YouTube asks for

YouTube's help page on custom thumbnails (read 2026-10-09) says a thumbnail should be an image format such as JPG or PNG, and it recommends a 16:9 aspect ratio for videos. Size limits depend on the device: 2 MB for video thumbnails on mobile and 50 MB on desktop. For Shorts it recommends 9:16, and for podcast playlists 1:1.

So the practical rule is to keep a 16:9 file under 2 MB if you upload from a phone. A 4K PNG can exceed that, so export JPEG if file size becomes a problem.

The test budget

A variant test is cheap because you only keep the winner. The table shows the three tiers; pick the tier by how large the thumbnail will be shown, not by habit.

Which tier to use

A thumbnail is shown small, often at a few hundred pixels wide. For most tests 1K is enough, and 15 images cost $1.50. Use 2K for the final pick if you want crisp text on a large TV or desktop view; one 2K image is $0.15. Use 4K only if you plan to crop into the image, since a 4K file is more likely to exceed the 2 MB mobile limit and to need a slower job path.

A cost-aware plan is to test all 15 at 1K ($1.50), then render the 5 winners at 2K (5 x 0.15 = $0.75), for $2.25 in total, the same as 15 images at 2K.

Reading the test

Keep the three variants for each video different in one thing only, such as the text, the face crop or the colour. If all three differ in everything, the result tells you nothing about which change mattered. Name the files with the video id and the variant letter so that the results map back to the plan.

Most thumbnail tools compare click-through over a window of days. The image side is the cheap part: the full 15-image set at 2K costs less than a coffee, so the real cost is the viewing time it takes to see a difference.

Generate the three variants for a video, pick one by whatever test you run, and keep the others as a record. If the first request returns a 202 instead of a 200, the generation is still running, and you read it from the job endpoints (see Jobs and results). Failed generations are not billed.

Keep text short and large in the prompt, and check the result at small size, since a thumbnail is often seen at a fraction of its pixels. Put any exact title text on in your own tool if spelling must be perfect.