20 hooks x 4 cuts = 80 vertical ad clips: cost by Sume model

Twenty hooks cut four ways is 80 clips, and at 8 seconds, 720p and 9:16 that test costs $48.00 on minimax-h3 (768p), $80.00 on wan-3.0 or gemini-omni-flash-1.1, $89.60 on kling-3 with audio off, $121.60 on seedance-2-mini and $193.60 on seedance-2-fast.

An ad test is a good first use of a replacement model after the Sora removal. OpenAI removed the Videos API and sora-2 on 2026-09-24 and listed no replacement, and an 80-clip test is small enough to run on several models before you commit a whole library.

Cost of the 80 clips

Each figure is the per-clip billable cents times 80. The hooks do not change the price; the model, length and resolution do.

What the platforms ask for

TikTok's ad specifications recommend 9:16 for non-Spark ads, with a minimum of 540x960, and accept mp4, mov, mpeg, 3gp and avi files up to 500 MB. YouTube Shorts accept square or vertical video up to three minutes. Check the pixel dimensions of the file you download against the platform minimum before you upload it, rather than assuming them from the resolution label.

The 8-second length is a choice, not a limit of either platform. If your best Sora ad was 6 seconds, kling-3 at 6 seconds costs 84 cents silent and wan-3.0 costs 75.

Naming the variants so results map back

Name each clip by hook number and cut letter, such as h07-b, and carry that name in your own table next to the job id. When the platform reports that h07-b won, you can open the prompt, the model and the cost in one row. Without that, an 80-clip test turns into 80 files with no lineage.

Keep the model id in the row as well. If you later re-render the winner on a different model, the new clip is a new variant, not a replacement for the tested one, and its performance has to be measured again.

Running the test

Give every clip a distinct idempotency key built from hook, cut and model, so a re-run of the script cannot double-bill. Submit in waves of 20 so that you stay inside the queue capacity of your workspace; the generation-admission docs describe 429 queue_full and say to retry with the same key when capacity opens.

Pick two models for the first pass, one cheap and one premium, and run 20 clips on each. At 20 clips the cost of the experiment is small: 20 x 100 cents = $20.00 on wan-3.0 and 20 x 242 cents = $48.40 on seedance-2-fast. Score on hook retention in your own analytics, not on how the clip looks in a preview.

• Keep hooks and cuts in a table, so a winner maps back to a prompt row.
• Render silent on Kling when an editor adds the voice later.
• Check platform policies on AI-generated content before you spend on a full run.