AI video price per second, Oct 2026: Seedance, Wan, Kling, H3 on Sume

On Sume in October 2026, the cheapest per-second price among the new video models is $0.0625 for Wan 3.0 or MiniMax H3 at 480p, and the dearest common one is Seedance 2.5 at 1080p, about $1.4216 a second at 16:9. Everything below is the provider list price times 1.25, which is how Sume bills every video row. Clip totals round up to the cent.

List prices and limits come from the Video Router docs and the video generation docs, read on 2026-10-06; the model launch dates behind the question are in Magic Hour's tracker. Provider list prices can change, so the live GET /v1/videos/models response is the number to trust before a large batch.

Per second at every resolution

Wan, Kling and H3 are priced per second. Seedance is priced per video token, so its per-second figure depends on frame size; the figures here are at 16:9.

What 5, 10 and 15 seconds cost

Totals at the cheapest and the most common resolution. Kling does not offer 480p, and the price it shows is the same at 720p and 1080p.

Why Seedance's price moves with the frame

Seedance 2.5 is billed per 1,000 video tokens, where tokens are width times height times seconds times 24 frames, divided by 1,024. At 1280 by 720 that is 21,600 tokens a second, and at 1920 by 1080 it is 48,600. List rates are $0.0214 per 1,000 tokens, and $0.0234 at 1080p. A vertical 720 by 1280 clip has the same pixel count as the horizontal one, so it costs the same; a 21:9 frame at 720p is 1470 by 630, about the same pixel count, so it lands within a percent of the 16:9 price.

This is why the Seedance row in the table says 16:9. If your ratio is different, the catalog's pricing_skus field gives the token rate and you can compute the figure yourself.

Reading the table for a budget

For 100 hook clips of 5 seconds, the 720p or 768p column gives $63.00 on Wan 3.0, $38.00 on MiniMax H3, $70.00 on Kling 3 without sound, and $289.00 on Seedance 2.5. Seedance 2.5 costs about 4.6 times Wan 3.0 and 7.6 times MiniMax H3 here, so the first decision in a large batch is the model, and the second is the resolution. Drafting at 480p and rendering only the winners at the final size cuts most of the spend, because the draft price is a fraction of the final.

A second cost lever is audio. Kling's sound adds 50 percent per second. H3's stereo is included in its rate, so there is no extra line. And one more is references: Seedance bills a reference video as extra input seconds at a discount, and MiniMax H3 charges extra after five reference images, so check the constraints for the id before a batch that uses them.

import math

RATE = {  # billed $ per second at 720p / 768p, list x 1.25
    "wan-3.0": 0.125,
    "minimax-h3": 0.075,
    "kling-3-audio-off": 0.14,
    "kling-3-audio-on": 0.21,
}
for model, rate in RATE.items():
    for secs in (5, 10, 15):
        total = math.ceil(rate * secs * 100) / 100
        print(f"{model:<18} {secs:>2} s  ${total:.2f}")

What the table leaves out

Three things are not in a per-second figure. Failed jobs are refunded, so a retry costs only the successful take. Reference inputs can add cost: MiniMax H3 includes five reference images and charges extra for each one after that, and Seedance prices a reference video as extra input seconds. And resolution upscales are separate rows: H3's 2K and 4K are priced only if you request them. None of these change the base rates above, but each can move a total, so the live catalog and the id's constraint text are worth reading before a batch that uses references.

A simple way to choose

If sound must be off and price matters most, Wan 3.0 or MiniMax H3 at 480p are the floor. If you want 15 seconds with sound and a clean 1080p frame, MiniMax H3 Max at 1080p is $3.00 and Kling 3 with audio is $3.15. If you want 30 seconds in one job, only Wan 3.0 and Seedance 2.5 qualify for ordinary text-to-video, and Wan is the cheaper of the two by a wide margin. The quality difference between them is the one thing a price table cannot show, so run one short test of each on your own prompt first.