Seedance 2.5 price by hand: video tokens and Sume's 1.25 margin

You can price a Seedance 2.5 clip on Sume with one line of arithmetic: tokens = width x height x seconds x 24 / 1024, multiplied by the provider rate of $0.0214 per 1,000 tokens ($0.0234 at 1080p), then by Sume's 1.25. A 10 second 720p 16:9 clip is 1280 x 720 x 10 x 24 / 1024 = 216,000 tokens, which is $5.78 billed.

Kling 4.0 was announced on 2026-09-28 with a 30 second ceiling, and the fal explainer (read 2026-10-05) says the full model is due in October. If you want a 30 second shot today, seedance-2.5 on Sume covers 4 to 30 seconds, so a budget formula is worth having before you queue anything.

The formula and its two catches

The estimator in packages/provider-pricing does not use a per-second price for Seedance 2.5. It converts the requested resolution and aspect ratio into a pixel size, multiplies by seconds and 24 frames per second, and divides by 1024 to get video tokens. The provider list price is per 1,000 tokens. Sume bills the provider list times 1.25 and reserves that amount at submit, as the video docs describe.

Two details change the answer. The rate steps up from $0.0214 to $0.0234 per 1,000 tokens at 1080p. And the pixel size is not exactly the nominal one: 21:9 at 720p is 1470 x 630, and 4:3 is 1112 x 834, so ratios other than 16:9 are slightly different in price. At 10 seconds the gap is only a few cents, but across a 30 second clip it reaches about nine cents at 720p, and it grows with resolution: 21:9 at 1080p is $42.85 for 30 seconds against $42.65 for 16:9. Square 1:1 is 960 x 960, which has the same pixel count as 16:9 at 720p, so the two cost the same.

Pixel sizes and a 10 second price for each

Scaling to other lengths

Because price is linear in seconds at a fixed size, you can scale from any row. A 30 second clip is three times the 10 second figure: $17.33 at 720p 16:9, $8.06 at 480p, and $42.65 at 1080p. A reference video changes this, as covered in the reference video break-even post.

Numbers like these are estimates for planning. The amount that counts is usage.cost on the finished job, and the models endpoint returns the supported durations, resolutions and ratios for each model. Treat the endpoint as the source for what is allowed and the job as the source for what was billed. If a provider changes its rate, the numbers here will move with the pricing package, so rerun the arithmetic instead of copying a figure from a month-old note.

The formula in Python

This script is the same arithmetic. It prints the 10 s and 30 s 720p figures from the table above.

SIZES = {
    ("480p", "16:9"): (864, 496),
    ("720p", "16:9"): (1280, 720),
    ("720p", "21:9"): (1470, 630),
    ("1080p", "16:9"): (1920, 1080),
}

def sume_price(resolution, ratio, seconds):
    w, h = SIZES[(resolution, ratio)]
    tokens = w * h * seconds * 24 / 1024
    rate = 0.0234 if resolution == "1080p" else 0.0214
    return round(tokens / 1000 * rate * 1.25, 2)

for seconds in (10, 30):
    print(seconds, sume_price("720p", "16:9", seconds))
print(sume_price("720p", "21:9", 30))

Using it before you submit

Use the formula to size a draft run. A 480p 16:9 pass costs $2.69 for 10 seconds against $5.78 at 720p, so a draft ladder at 480p saves more than half before you commit to a final. The video generation docs list the sizes each model accepts, and the Video Router docs describe the older flat fields.

The same arithmetic explains why resolution is the biggest lever. Going from 720p to 1080p multiplies the token count by 2.25 and raises the rate by about nine percent, so a 10 second clip goes from $5.78 to $14.22, which is nearly two and a half times. Going down to 480p cuts the cost to under half. For a shot you will regenerate several times, pick the lowest resolution that lets you judge motion.

Before you submit a batch, add the reserved amount for every job and compare it with your balance. Submitting with an Idempotency-Key per shot keeps a retried request from reserving twice.