$25 of 8-second AI video on Sume: how many clips per model

$25 buys 25 eight-second clips on Gemini Omni Flash 1.1 at 720p ($1.00 each) and the same 25 on Wan 3.0 at 720p. It buys 31 on MiniMax H3 Max at 768p ($0.80), 16 on Omni at 1080p ($1.50), and 83 on Omni at 360p ($0.30). These come from the Video Router catalog's list prices times 1.25, rounded up to the cent per job (read 2026-10-06). The Sora API they replace ended on 2026-09-24 (Magic Hour tracker, read 2026-10-06).

Eight seconds is a useful unit: it fits Omni (3-10 s), Wan (2-30 s), H3 and H3 Max (5-15 s), so one test brief can run on all of them. Seedance models are priced per video token and are left out of the count.

Clips per $25

The arithmetic is the same for every row. Take the provider list price per second from the Video Router model table, multiply by 8 seconds and by 1.25, round up to the cent, then divide $25 by that figure and drop the remainder. For Gemini Omni Flash 1.1 at 720p the list rate is $0.10 per second, so 8 seconds is $0.80 at list and $1.00 billed. Wan 3.0 is listed at $0.05, $0.10 and $0.20 per second by resolution (fal, 2026-08-24, per the Sume catalog notes), and MiniMax H3 at $0.05 and $0.06 per second for 480p and 768p. H3 Max is $0.05, $0.08 and $0.16 for 480p, 768p and 1080p. Each of those multiplies by 1.25 on Sume, the same house margin on every row.

Two things stand out. Wan 3.0 and Omni cost the same at 720p ($1.00 for 8 seconds), so the choice there is about output quality and supported inputs, not price. And the resolution tier moves cost more than the model does: Omni at 1080p is 50 percent dearer than at 720p and five times the 360p price, so decide the tier before you decide the model.

What the table does not say

• Quality differs by model. A cheaper clip that you re-run three times costs more than one good clip, so budget for retakes before you compare tiers.
• Vertical output is documented for Omni (16:9 and 9:16 only) and for Seedance 2.0. For Wan and H3, read supported_aspect_ratios before you count vertical clips.
• H3 at 768p is native, and the catalog says 720p is not a mode for it. H3 Max's 1080p is a latent refinement from native 768p.
• Reserve and balance: Sume reserves the billable amount at submit, so a balance under one clip's price returns an insufficient-credits error before any work starts.

Count clips from your own balance

Keep the arithmetic in one function so a dashboard and a submit guard agree. The prices below are the per-8-second billable values from the table.

PRICE_8S = {
    "omni-360p": 0.30,
    "omni-720p": 1.00,
    "omni-1080p": 1.50,
    "wan-480p": 0.50,
    "wan-720p": 1.00,
    "h3-max-768p": 0.80,
}


def clips_for(budget_usd: float) -> dict:
    cents = round(budget_usd * 100)
    return {k: cents // round(v * 100) for k, v in PRICE_8S.items()}


print(clips_for(25))

A sensible spend plan

Split the $25 in two. Spend about $5 on cheap drafts: sixteen Omni 360p clips at $0.30 cost $4.80 and prove the brief and the framing. Spend the remaining $20 on finals at the tier you will ship, which is twenty Omni 720p clips at $1.00, or thirteen at 1080p ($19.50, with $0.50 spare). This is the same draft-then-final habit that image teams use, and it is cheaper than discovering a bad prompt at the highest tier.

If you want a three-model comparison before you commit, the smoke test post runs one prompt through three models in Python. For historical context on what the retired API cost, see the per-second history.

Cost per keeper, not per clip

The table counts attempts. What you ship is keepers, and the gap between the two is your retake rate. If one 8-second Omni 720p clip in three is good enough to publish, each keeper costs three attempts, or $3.00, and $25 yields eight keepers rather than twenty-five. At 360p the same one-in-three rate costs $0.90 per keeper, which is why drafting low is not a saving on paper only.

Measure your own rate before you trust anyone's. Run twenty prompts at the cheapest tier, mark the ones you would publish, and divide. A brief that is clear to the model has a high rate on any model; a vague brief has a low rate on every model, and moving to a pricier tier does not fix it. Spend the first few dollars on fixing the brief.

• Keep a ledger of attempts and keepers per model; one spreadsheet column each is enough.
• Do not add the reservation to the total on top of the finished cost: the hold is released into the captured amount, and usage.cost is the figure to sum.
• A job still reserves at submit, so a $25 balance with $24 already in flight can refuse a new submit with a 402 even though the jobs have not finished.