Test 5 avatar ad hooks, then scale the winner: $12.87 on Sume

A hook test on Sume Avatar 1.0 costs $5.52 for five 6-second clips at standard ($0.184 per second), and scaling the winning script to a 30-second plus ad adds $7.35 ($0.245 per second). The whole loop is $12.87 with an avatar you already have, or $13.82 if you create one first at $0.95.

The launch of Vidu Q4 Preview on 2026-10-07 put cheap per-second video back in the testing conversation. Unite.AI's coverage lists a $0.014 per second headline price, optional synced dialogue, and up to 16-second clips (read 2026-10-12). Sume does not ship Vidu, and a generic video model is not the same product as a scripted talking avatar. If your test is about what a presenter says in the first six seconds, the unit that matters is the exact script, spoken by the same face every time.

What does each step cost?

Avatar video is billed per second by tier. standard is the fastest path, plus is the default when you omit quality, and max is the highest tier. Sume's docs accept a script that it estimates at 4 to 60 seconds, so a 6-second hook is valid and a 3-second one is not.

Add the rows you actually use. Test and scale on plus with a product image and you pay $13.56; add the avatar and it is $14.51. Finishing the winner on max instead of plus turns $12.87 into $22.02.

How do I run the five hooks?

Create the avatar once with a stable avatar_handle, then submit five POST /v1/avatar-1.0/talking-video jobs that differ only in script. Give every request its own Idempotency-Key (for example hook-a to hook-e) so a retry returns the same job instead of billing again.

Set the same aspect_ratio (the default is 9:16) and the same scene prompt on all five. If the background changes with the script, you are testing the background. Sume's Generate avatar video page lists 720p as the only resolution today, which is fine for a test.

Will five jobs run at once?

That depends on your plan. Sume's Generation admission page lists processing concurrency of 1 on Free, 4 on Pro, 8 on Startup, and 20 on Scale. On Pro, four hooks process and the fifth waits as queued, which is not a failure. Store each job_id and poll.

How do I pick a winner and scale it?

The cost shape is the point. Testing is about 43 percent of the whole $12.87 loop, and you spend the larger tier only on a script that has already earned it.

• Decide the metric before you spend: hold rate at 3 seconds, click-through, or cost per view.
• Run the hooks as separate ads to the same audience, equal budget, same window.
• Re-render only the winner as a 30-second script that opens on the winning hook verbatim. The script must be new text, so the full ad is a new job, not an extension.
• Keep the loser scripts. A hook that lost with one avatar may win with another handle.