Cost per kept second: Omni 720p vs H3 Max 768p at 1 take in 3

A generated second only counts when you keep the take. On Sume, Omni Flash 1.1 at 720p bills $0.125 per second and MiniMax H3 Max at 768p bills $0.10, so if you keep one take in three, a kept second costs $0.375 on Omni and $0.30 on H3 Max. The cheaper rate stops winning only when its model needs 25% more takes than the other to give you the same keeper.

Failed or rejected takes are not what this post is about. It prices the takes you discard because the motion, the face or the product looked wrong. Every rate below comes from the Sume price tool (read 2026-10-09) and every cost is a simple multiplication of it.

The kept-second table

Cost per kept second is the per-second rate times the number of takes you generate for each keeper. The columns assume you generate 1, 3 or 5 takes per keeper and bill every one of them at full length.

What an 8-second keeper costs

For a single 8-second keeper at one keeper in three takes, Omni at 720p costs 3 x $1.00 = $3.00, and H3 Max at 768p costs 3 x $0.80 = $2.40. The gap is $0.60 per keeper, so it only matters when you ship dozens of clips.

The break-even is the ratio of the two rates. H3 Max at 768p ($0.10) against Omni at 720p ($0.125) means H3 Max needs to stay under 1.25 times Omni's takes per keeper. If Omni gives you a keeper in 3 takes, H3 Max must give you one in 3.75 takes or fewer to stay ahead. Count your own rejects over a first batch of ten before you decide; this post gives no keep-rate for either model.

Read the rates from the catalog

Do not hard-code the table. GET /v1/videos/models returns pricing_skus for each model, with keys of the form per-video-second-<resolution> for per-second rows. The script below builds the same cost per kept second from live data.

import requests, os
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
r = requests.get("https://api.sume.com/v1/videos/models", headers=H, timeout=30)
r.raise_for_status()
rate = {}
for model in r.json()["data"]:
    for sku, usd in model["pricing_skus"].items():
        if sku.startswith("per-video-second-"):
            rate[(model["id"], sku.removeprefix("per-video-second-"))] = float(usd)

def per_kept_second(model, res, takes_per_keeper):
    return rate[(model, res)] * takes_per_keeper

for model, res in [("gemini-omni-flash-1.1", "720p"), ("minimax-h3-max", "768p")]:
    print(model, res, round(per_kept_second(model, res, 3), 4))

Scale it to a batch of keepers

The same rule scales linearly. Ten keepers of 8 seconds at three takes each means 30 takes. Multiply the per-take price by the take count, and compare the models on that total rather than on the headline rate.

Limits and gotchas

The Video generation docs list the supported durations and resolutions per model, and the Video Router page lists the request fields.

• Omni Flash 1.1 accepts 3-10 seconds; H3 Max accepts 5-15 seconds. A keeper longer than 10 seconds cannot come from Omni in one take.
• H3 Max has no 720p row: send 768p. Omni has no 768p row.
• Native audio is always on for both models, so there is no cheaper silent rate to compare against.
• Sume reserves the full price when you submit, so a discarded take is still a full-length charge. Cut the cost of a bad take with a shorter test length or a lower resolution.