Seedance 2.5 reference video: price flips at 22.5 seconds on Sume

Under 22.5 seconds, a reference video makes a Seedance 2.5 job cost more on Sume; above 22.5 seconds it costs less. At 720p and 16:9, a 10 s clip rises from $5.78 to $8.67 when you attach reference_video_urls, while a 30 s clip falls from $17.33 to $15.60. The reason is in the price estimator in packages/provider-pricing, which treats a reference video as 15 extra billable seconds and then multiplies the total by 0.6.

This matters this week because Kling 4.0 is being sold on long clips and many references. The fal explainer for Kling 4.0 (read 2026-10-05) lists up to 5 reference videos with 30 seconds combined, and a 30-second ceiling for the full model. The closest Sume match for a long clip with a video reference is seedance-2.5, which accepts 4 to 30 seconds and lists image, video and audio references.

What the estimator does with a reference video

The estimator checks one thing about reference videos: whether the request has at least one. It does not measure their length or count them. When one is present it assumes 15 seconds of input, adds that to the output seconds for the token count, prices the sum at the model's per-token rate, then applies a 0.6 multiplier. Sume then applies its list times 1.25 margin, as for every catalog model.

So the billable shape is 0.6 x (D + 15) / D times the plain price for a D-second clip. That equals 1 when D is 22.5. Below that, the 15 assumed seconds dominate; above it, the 0.6 discount wins. Whole-second durations make the crossover land between 22 and 23 seconds.

Price by length, with and without a reference video

How to read the table

These are estimates computed from the pricing package, the same numbers Sume uses to size the hold at submit. The Seedance 2.5 docs describe billing as provider list times 1.25 with a reserve at submit, and usage.cost on the finished job is the amount Sume bills. Check the final figure on the job rather than trusting a table, because the provider rate card can change. If you budget a batch, compute the hold for every shot first, then compare it with your balance before you submit, so a late shot is not refused for lack of credit.

Two practical readings follow. First, a short clip that uses a reference video for identity or motion should be priced at the higher figure, not at the plain per-second rate. Second, if you were going to render a 24 to 30 second clip anyway, a video reference does not raise the bill and slightly lowers it.

Reproduce it in Python

The formula is short enough to check yourself. This runs as written and reproduces the 30 s row.

def seedance_25_720p(seconds, reference_video=False):
    billable = seconds + (15 if reference_video else 0)
    tokens = 1280 * 720 * billable * 24 / 1024
    provider = tokens / 1000 * 0.0214
    if reference_video:
        provider *= 0.6
    return round(provider * 1.25, 2)

for s in (10, 22, 23, 30):
    print(s, seedance_25_720p(s), seedance_25_720p(s, True))

When it is worth it

Use the video reference for what it is for, not for a discount. It carries motion or identity from a clip you already have, and seedance-2.5, seedance-2, wan-3.0, minimax-h3 and minimax-h3-max accept video references, while gemini-omni-flash-1.1 takes video references of up to 3 seconds each. Read supported_input_references from GET /v1/videos/models before you send one.

For a clip over 22 seconds, add the reference only if it helps the result. For a 5 to 10 second shot, expect the higher price and consider an image reference first, since images do not trigger the 15 second assumption in the estimator. For a long clip with a reference video, run a 480p draft first to check motion, then render the final at 720p or 1080p. Send an Idempotency-Key on every submit so a retry does not create a second job; the video generation docs list the field, and the jobs docs cover polling.