$5 prompt-testing budget on Wan 3.0: how many 2-second drafts

On Sume, $5 buys 40 two-second Wan 3.0 drafts at 480p, 20 at 720p or 10 at 1080p. Wan 3.0's 2-second floor makes it the cheapest way to test whether a prompt works at all (read 2026-10-09).

Drafts per $5

Prompt testing is about breadth: ten wordings, four camera moves, three lighting setups. A 2-second clip answers 'does the model understand this scene?' at the lowest price. The 4-second column shows the cost of seeing the motion develop.

Run a grid

Give each prompt its own idempotency key so a rerun of the script never pays twice. This submits the same prompt in several camera moves.

import os, requests

H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
moves = ['slow push-in', 'orbit left', 'static wide', 'handheld follow']
for i, move in enumerate(moves):
    body = {'model': 'wan-3.0', 'duration': 2, 'resolution': '480p',
            'prompt': f'A red kite over a beach at dusk, {move}'}
    r = requests.post('https://api.sume.com/v1/videos',
                      headers={**H, 'Idempotency-Key': f'kite-grid-{i}'},
                      json=body, timeout=60)
    print(i, r.status_code, r.json().get('id'))

Spending the five dollars

A grid of 8 prompts x 4 camera moves is 32 drafts. At 2 seconds and 480p that is 32 x $0.125 = $4.00, leaving $1.00 of the $5.

Upgrading the three best to 5 seconds at 720p costs $1.875, and the single winner at 10 seconds in 1080p costs $2.50. The total spend would be $8.375.

That path goes from many cheap tests to one expensive keeper. It only works if the draft is a reliable predictor, so keep the prompt wording identical when you upgrade.

Limits

• A 2-second draft shows the opening of a scene, not the full motion. Lengthen the winner.
• Admission has queue limits per plan. A queue_full response means wait or cancel.
• No v1 model accepts a seed, so you cannot upscale a draft; the final is a new generation.
• The $5 here is customer USD. The balance reservation at submit equals the price shown.