A/B test 6 hooks, 2 openings each: 12 vertical 5-second clips priced

Six hooks with two different openings each is 12 test clips. At 5 seconds, 9:16, the round costs $4.50 on MiniMax H3 (768p), $7.50 on Wan 3.0 (720p), $22.68 on Seedance 2 and $34.668 on Seedance 2.5.

The test only has to be cheap enough that you run it every week. The last column adds the cost of finishing the two winners at 1080p on Wan 3.0, which is the realistic total for one round that ends in two shippable ads.

Cost of one test round

Each clip is a 5-second job priced by the tool. The test-round column is 12 x the clip price. The final column adds 2 x the Wan 3.0 1080p 5-second price, $1.25 each.

Design the test so the price buys a signal

Change one thing between the two openings of a hook, such as the first-frame subject or the on-screen question, and keep the length identical. The model is not a variable in a hook test: pick one for the test, because Seedance, Wan and MiniMax look different and the audience may react to the look.

Models differ in their sound handling. Gemini Omni Flash 1.1 always generates synced audio, so a test of spoken hooks belongs there; Wan 3.0 hooks are better judged silent with captions.

Screen at 480p, finish at 1080p

If the test is only about the idea, run it at 480p: 12 Wan 3.0 clips at 5 seconds cost $3.75 ($0.3125 each). The two winners then cost $2.50 at 1080p, for a round total of $6.25.

Wan 3.0 is a flat per-second rate, so this arithmetic is exact. Token-priced models such as Seedance depend on resolution, seconds and aspect ratio, so price them with the estimate for the exact request. Remember the final is a new generation, not an upscale: no model in the v1 catalog accepts a seed, so a winning draft cannot be reproduced at higher resolution.

Submit the round with one key per cell

Name each idempotency key by hook and variant, so a retry of the same cell returns the original job and a re-roll gets a new key.

import os, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
hooks = ["price", "problem", "proof", "question", "story", "offer"]

for h in hooks:
    for v in ("a", "b"):
        body = {"model": "wan-3.0", "prompt": f"Vertical hook, {h}, opening {v}",
                "resolution": "720p", "duration": 5, "aspect_ratio": "9:16"}
        r = requests.post("https://api.sume.com/v1/videos", json=body, timeout=60,
                          headers={**H, "Idempotency-Key": f"hooktest-{h}-{v}"})
        print(h, v, r.status_code)

Gotchas

• 12 jobs can exceed the processing slots of a small plan: extra jobs wait as queued, and only a full queue returns 429 queue_full.
• Seedance models take no seed, so two openings are two different generations by design.
• Check each model's supported_durations: MiniMax H3 starts at 5 seconds, Seedance 2.5 at 4, Wan 3.0 at 2.