A $25 test budget across five new video models: 8-second clip counts

A $25 test budget on Sume buys about 5 eight-second 720p clips on seedance-2.5, 25 on wan-3.0, 14 on kling-3 with sound at 1080p, 41 on minimax-h3 at 768p and 25 on gemini-omni-flash-1.1 at 720p. Prices are billed rates with the 1.25 multiple, and failed jobs are refunded, so the counts are for clips that rendered.

Seedance 2.5 and MiniMax H3 are both dated 2026-07-31 on Magic Hour's tracker, read 2026-10-06, and Kling 3.0 is dated 2026-02-05. Limits are in the Video Router docs and prices in the video docs.

Clip counts for $25

An 8-second 16:9 clip fits the length range of every model on the list. Counts are rounded down. Seedance 2.5 is billed by video tokens, so its price moves most with resolution: the same 8-second clip at 1080p is about $11.38, which is why the table stops at 720p.

How to spend it

Do not spend evenly. Put a third on cheap drafts, where you vary the prompt, and the rest on the finalists at the resolution you would ship. For example, run ten prompts at Wan 480p for $5.00, pick two, then run each of those on Seedance 720p, Kling 1080p and H3 768p, which is six clips for $13.82. That is a full comparison for $18.82, which leaves $6.18 for retakes.

A spend guard in Python

This keeps a running total from the same rates, so a loop stops before it crosses the cap. Prices come from the pricing tables above; update them if the catalog changes.

RATES = {"wan-3.0@720p": 0.125, "minimax-h3@768p": 0.075,
         "kling-3@1080p+sound": 0.21, "gemini-omni-flash-1.1@720p": 0.125}
CAP = 25.0

def plan(jobs):
    spent, ok = 0.0, []
    for key, seconds in jobs:
        cost = RATES[key] * seconds
        if spent + cost > CAP:
            break
        spent += cost
        ok.append((key, seconds))
    return ok, round(spent, 2)

jobs = [("wan-3.0@720p", 8)] * 30 + [("kling-3@1080p+sound", 8)] * 5
print(plan(jobs))

The guard is a planning aid, not a limit on your account. The real control is the credit balance: a submit with too little returns 402 insufficient_credits, and the price is reserved when you submit, so you cannot overshoot silently. For unit costs by the minute, the cost per minute post lays out the catalog, and the by-length mapping helps you pick which model gets which prompt. Before any of it, check the request against the catalog.

What the counts do not tell you

A clip count is not a quality count. Wan 3.0 buys the most clips, but if two of every three miss, the cost per kept clip rises. Track kept clips per dollar, not rendered clips per dollar. For a $25 budget, a model that keeps one in two at $4.63 a clip costs $9.26 per keeper, and a model that keeps one in four at $1.00 costs $4.00.

The test is cheapest when the prompts are specific. A vague prompt wastes clips on the model guessing. Write the subject, the camera move, the light and the sound in one or two sentences, and keep it the same across models.

Finally, spend the last few dollars on the one test that matters most to you. If your product needs a person to speak, test that. If it needs a logo to stay readable, test that. A narrow, honest test of your own hardest shot tells you more than a broad test of easy ones.

Write the result down in a table with the prompt, the model, the setting, the cost and a keep-or-drop mark. After thirty clips you will see which model fits which kind of shot, and that table is your buying decision, with prices you have actually paid rather than prices you read.