Cost per kept AI video clip when 1 in 2, 3 or 4 is a keeper on Sume

The cost of a usable 10-second clip is the clip price divided by your keep rate. If one clip in two is a keeper, the cost per kept clip is twice the sticker price: $2.50 on Wan 3.0 at 720p, against $1.25 a clip. At one in four it is four times: $5.00. A model that costs more per clip can still cost less per kept clip if it needs fewer tries.

The prices below are Sume's Video Router rates for 10-second clips, the provider list rate times 1.25 and rounded up to the cent per job (read 2026-10-06). The keep rates are not benchmarks. They are the four numbers you are most likely to see in your own review, so replace them with yours.

Break-even between a cheap and an expensive model

Divide the two prices to find the break-even keep rate. Wan 3.0 at 720p is $1.25 and Seedance 2.5 at 720p in 9:16 is $5.78, a ratio of 4.6. If Wan keeps one clip in four and Seedance keeps one in two, Wan costs $5.00 per kept clip and Seedance $11.56. Seedance needs a keep rate 4.6 times Wan's before it wins on cost alone, and with a keep rate of one in two against Wan's one in four it falls well short.

That does not make the cheap model the answer for everything. Cost per kept clip ignores that a model may be unable to make the shot, so it cannot replace a look test. What it does is turn an argument about taste into a number that a producer can set a budget on.

• Count a clip as kept only when it would ship, not when it is close.
• Count retakes you submit by hand and the ones your pipeline submits for you.
• Failed jobs are refunded, so they do not enter the keep rate. Rejected clips do.
• Draft on a cheaper tier and finish on the tier you ship, so the expensive model sees only briefs that already work.

Rank models from your own review log

Keep a log of one row per rendered clip: model, price, kept or not. The script below prints cost per kept clip for each model in the log, so the ranking updates as you add rows.

from collections import defaultdict

LOG = [
    ("wan-720p", 1.25, True), ("wan-720p", 1.25, False),
    ("wan-720p", 1.25, False), ("wan-720p", 1.25, False),
    ("seedance-720p", 5.78, True), ("seedance-720p", 5.78, False),
]

spend = defaultdict(float)
kept = defaultdict(int)
for model, price, ok in LOG:
    spend[model] += price
    kept[model] += ok

for model in spend:
    per_kept = spend[model] / kept[model] if kept[model] else float("inf")
    print(model, round(per_kept, 2))

A worked example at 1080p

Take a brief that needs 1080p. Gemini Omni Flash 1.1 at 1080p is $1.88 for 10 seconds, MiniMax H3 Max at 1080p is $2.00 and Kling 3 with sound is $2.10. Suppose the review keeps one clip in three on Omni, one in two on H3 Max and one in two on Kling. The cost per kept clip is then $5.64, $4.00 and $4.20. The cheapest sticker is the most expensive keeper.

The example uses made-up keep rates on purpose. The point is the shape of the calculation: the price column from the catalog does not rank models until the keep rate is in it. Run the same brief through each model five times before you decide, and write the keepers down.

Budget for the average, then add a margin

Take the keep rate from the last 20 clips and budget 20 percent more clips than you plan to ship. Each job reserves its price at submit, so the balance cannot be overspent, but it can run out in the middle of a batch, so a small margin means you finish the campaign instead of explaining a 402 insufficient_credits. The model prices are in the Video Router catalog.