AI b-roll instead of stock footage: 50 six-second clips cost

If you want generic b-roll (coffee being poured, rain on a window, hands on a keyboard) without searching a stock site, generate it. Fifty 6-second clips cost $11.50 on Gemini Omni Flash 1.1 at 360p, $19.00 on Wan 3.0 or MiniMax H3 Max at 480p, and $37.50 on Omni Flash or Wan 3.0 at 720p. Sume bills the list price times 1.25, rounded up to the cent for each clip.

These are generation costs only. They say nothing about whether a generated shot is as good as a filmed one; for that you have to look at the output, which is why the cheapest tier is a sensible first pass.

Price of one 6-second clip, by model and tier

Each clip is rounded up to the next cent, so 50 clips are 50 times the rounded clip price. The lists below were read from the fal and Sume pages today.

Pick the tier by where the clip will run

A 360p clip is a proof of concept. Use it to test whether a prompt produces the composition you want, then re-render the keepers at 720p or 1080p. Omni Flash is the cheapest way to do that because its 360p tier is $0.03 per second on the vendor page, the lowest rate in the table.

For a website hero loop or a social background, 720p is usually enough. Reserve 1080p and above for shots that fill the whole frame on a large display. Omni Flash 1.1 accepts 360p, 720p, 1080p and 4K, and Wan 3.0 and H3 Max accept 480p, 720p or 768p, and 1080p.

Limits that change a b-roll plan

• Omni Flash clips run 3 to 10 seconds and only in 16:9 or 9:16.
• H3 Max has a 5-second minimum, so a 3-second cutaway must come from Omni, Wan 3.0 (2 s minimum) or a trim.
• Every Omni Flash and H3 Max clip comes with audio. Kling Video v3 Pro lets you switch audio off, which is the $0.14 row above.
• The API has no seed field, so you cannot regenerate the same take. Download what you like.

Queue a batch

This script submits three shots at 480p and prints the polling URLs. It reads the key from SUME_API_KEY and does not wait, so it is safe to run for a list of 50. Poll each URL later until the status is completed.

import os

import requests

API = "https://api.sume.com/v1/videos"
HEADERS = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
SHOTS = [
    "Slow push-in on a steaming espresso cup on a wooden counter, morning light",
    "Hands typing on a laptop in a bright office, shallow depth of field",
    "Rain on a city street at dusk, reflections on wet asphalt",
]

polling_urls = []
for prompt in SHOTS:
    body = {
        "model": "wan-3.0",
        "prompt": prompt,
        "duration": 6,
        "resolution": "480p",
        "aspect_ratio": "16:9",
    }
    resp = requests.post(API, headers=HEADERS, json=body, timeout=60)
    resp.raise_for_status()
    polling_urls.append(resp.json()["polling_url"])

print(len(polling_urls), "jobs queued; poll each polling_url until status is completed")

Budget check

A failed job is refunded, so the table is the worst case for a clean run. If you plan to keep one clip in three, divide the library price by the keep rate: at 480p on Wan 3.0, 50 kept clips cost about 150 generations, or $57.00. Decide your keep rate on a pilot batch of ten before you commit to a full library, and write the prompt list in a file so a rerun uses the same wording and you can compare tiers fairly.