Thumbnail A/B set per video: three models for 31 cents on Sume

A three-way thumbnail test for one video costs 31 cents on Sume: one 16:9 image from GPT Image 2.5 at high (7 cents), one from Nano Banana Pro at 1K (19 cents) and one from Seedream 5.0 Lite (5 cents). Ten videos cost $3.10. All three rows list 16:9, so the same request body works with only model changed.

The three rows

Prices are catalog list times 1.25, rounded up to the cent. The docs say Sume bills the endpoint pricing lines, which already include the margin, and that you pay cost_usd x n.

Running the test

Use the same prompt text, with the title or hook in quotes, on all three. Ask for 16:9 and leave n at 1 so each model gets one try; if you want more, n: 4 on Seedream costs 20 cents. Download the three URLs and put them side by side at 168 by 94 pixels, the size a thumbnail shrinks to in a sidebar, before you judge.

• Judge at small size first, then full size.
• Count legible words at small size; that is the real thumbnail test.
• Keep the winner's model id for that series so the channel looks consistent.
• Run the test again whenever you change the title style, not for every video.

Does the test pay for itself

If you publish ten videos a month, the test costs $3.10 a month. A single thumbnail swap on a video that earns a better click rate is worth more than that, and the test also tells you which model to use for the next batch. If you already know one of the three always wins for your style, drop the other two and the cost falls to the one-model price.

For a cheaper variant, replace Nano Banana Pro with Nano Banana 2.1 at 10 cents; the set becomes 22 cents.

Pitfalls

Text in a thumbnail is where models disagree most, and OpenAI's guide notes text rendering can still struggle with precise placement and clarity, so a clean render on one model is not proof for the other two. Do not trust sume/auto for a controlled test, since the docs say Sume never discloses which family it ran. Pin all three ids.

Finally, check usage.cost on each response and add it to your sheet. If the sum is not 31 cents, one of the rows changed price, and the catalog endpoint will show which.

Storing the results

Name each output file with the model id and quality, for example video42-gpt-image-2.5-high.png, and keep a single sheet with one row per video: the three costs from usage.cost, the winner and the click result after a week. After ten videos you have enough rows to see whether one model keeps winning for your channel, and the test can shrink to that model plus one challenger. Keep the prompt text in the same row so you can see which wording the winner came from.