Seedance 2.5 with a reference video: costs more under 22.5 s

Adding a reference video to a Seedance 2.5 request on Sume makes a short clip more expensive, not cheaper. The price is multiplied by 0.6, but the token count includes 15 assumed seconds of input video on top of your output seconds, so an 8 second 720p clip rises from $4.63 to $7.98 and only clips longer than 22.5 seconds come out cheaper.

This is easy to miss because the 0.6 multiplier looks like a discount. Both numbers come from Sume's Seedance estimator in the repository; the reference-video support itself is listed on Video Generation, which says the Seedance 2.x models accept audio and video references. Read 2026-10-10.

The rule in one line

Sume estimates cost as tokens for (output seconds + 15) at the model's per-token rate, then takes 60% of that. The 15 seconds is a fixed assumption for the length of the input video, not a measurement of your file. A 3 second reference and a 14 second reference reserve the same amount.

The break-even output length solves 0.6 x (d + 15) = d, which gives d = 22.5 seconds. Below it the reference costs more than a text-only request for the same output; above it, less. The longest Seedance 2.5 output is 30 seconds, where the reference version is cheaper by 10%.

Seedance 2.5 at 720p, 16:9

Prices are per job, list x 1.25, rounded up to the cent after the 0.6 step.

Planning around it

If you use a reference clip for motion or style on 5 to 10 second shots, budget the with-reference column, not 60% of the plain price. A sheet that applies a 0.6 factor to the text-only cost will under-reserve by about 75% at 5 seconds ($1.73 against $6.94) and the wallet check will return 402 before the batch finishes.

Where the reference is optional, a still image reference or first-frame image is a cheaper way to steer a short clip, because the 15 second assumption applies to video input only. For a long single shot of 25 to 30 seconds the video reference is the cheaper path, since the 0.6 factor wins.

One more caution: the 15 second figure is an assumption in the estimate, not a read of your upload. If a request fails after the provider starts, the hold is released, so the extra reserve is a temporary cost to your available balance rather than a charge. It still matters when several jobs run at once, because each reservation reduces what the next job can draw on.

Other rows treat references differently. Gemini Omni Flash 1.1 accepts video references but bills output seconds at its resolution rate, and Wan 3.0 bills per second by resolution, so neither adds an input-length term. That makes Wan 3.0 a more predictable choice when you cannot tell a reference's length ahead of time.

• Under 22.5 s of output: reference video costs more.
• Exactly 22.5 s: break-even.
• Over 22.5 s: reference video costs less, up to 10% at 30 s.

Checking before you send

Compute both columns before a batch and submit the cheaper one only if the creative result is equal. The reservation is made at submit; a short balance is rejected as 402 insufficient_credits with no provider work, and failed jobs release their hold, which you can see in the Usage ledger as reserved, captured and refunded rows.

A simple rule for a team sheet: add a column named "ref video" and set it to yes or no per shot. When it is yes and the output is under 22 seconds, use the with-reference price; when it is no, use the plain price. Do not blend them with a single discount, because the two lines cross only once, at 22.5 seconds, and most short-form work sits well to the left of that point.