Does a reference video cost more on Seedance 2.5? Break-even at 22.5 s

A reference video makes a short Seedance 2.5 clip cost more and a long one cost less. On Sume the crossover is 22.5 seconds: below it the reference clip is pricier, above it cheaper. At 720p a 4-second clip goes from $2.32 to $6.59, while a 30-second clip goes from $17.34 to $15.61.

The rule in the code

Seedance 2.5 is priced by tokens: width x height x seconds x 24 / 1024. When a request includes any reference_video_urls, the estimator in packages/provider-pricing/src/pricing-normalize.ts assumes 15 seconds of input video and bills (duration + 15) x 0.6 seconds instead of the duration. The rate itself is unchanged, and Sume still adds its 1.25 margin and rounds up to the cent.

Set the two expressions equal and the result is duration = (duration + 15) x 0.6, so duration = 22.5. The table shows both sides around that point. Reference images and audio do not trigger this rule; only reference videos do.

Billed price with and without a reference video

What to do with it

If you only need the style of a short clip, a reference image is cheaper than a reference video because it carries no input-seconds charge. Reserve reference videos for work where the motion has to match, and expect a 4-second clip to cost about 2.8 times a plain one at 720p.

For clips of 23 seconds or more the reference video is a small discount. Do not choose a longer clip just to chase it, since the 30-second reference clip still costs $15.61 at 720p. Seedance 2.5 accepts 4 to 30 seconds per the Sume video models docs, and the Video Router docs list the reference fields.

• Under 22.5 s: reference video costs more.
• Over 22.5 s: reference video costs less.
• The estimate assumes 15 s of input whatever your file's real length.

Check your own number

This sketch reproduces the 720x1280 rows. Pass the rate for the tier you use; 21,400 micro-dollars per 1,000 tokens is the 480p/720p Seedance 2.5 rate in pricing-tables.ts.

For a full script that handles four models, see the ceil-to-the-cent estimator.

import math

def seedance_25_cents(seconds, w=720, h=1280, ref_video=False, rate=21_400):
    billed = (seconds + 15) * 0.6 if ref_video else seconds
    tokens = w * h * billed * 24 / 1024
    provider = math.ceil(tokens * rate / 1000)
    return math.ceil(provider * 125 / 100 / 10_000)

for s in (4, 10, 22, 30):
    print(s, seedance_25_cents(s), seedance_25_cents(s, ref_video=True))

Budgeting a reference-video job

Add the reference-video line to your estimate before submitting, not after. A team planning ten 6-second remixes at 720p with a reference video would see about $7.28 each (6 s plus 15 assumed input seconds, times 0.6, is 12.6 billed seconds) against $3.47 without one, so the batch is $72.80 instead of $34.70. Only the estimate treats the input as 15 seconds, so a shorter or longer source file does not move the price either way.

Because the rule is part of the quote, you can always confirm it by comparing two estimates that differ only in reference_video_urls. If they do not follow the (duration + 15) x 0.6 pattern, trust the quote and ignore this page; the code is the authority and this article is a dated reading of it.