Seedance 2.5 reference video: 15 assumed input seconds at 60%

Seedance 2.5 accepts reference videos, and they change the price. Sume's pricing code does not know how long your reference clip is at estimate time, so it assumes 15 seconds of input, adds them to the output seconds for the token count, and then applies a 0.6 multiplier to the provider cost. Two constants in the pricing tables carry this: 15 assumed seconds and a 6/10 multiplier.

Worked example at 720p, 16:9

For a 10-second output with one reference video, billable tokens are for 25 seconds, not 10: 1280 x 720 x 25 x 24 / 1024 = 540,000. At $0.0214 per 1,000 that is $11.56, times 0.6 is $6.93, times the 1.25 Sume multiple is $8.67. The same clip without a reference video is $5.78.

What to take from it

The fal page states the plain per-token rule for text-to-video (read 2026-10-05); the reference-video handling is Sume's own estimating rule, so check GET /v1/video-router/models and your first completed job rather than extrapolating from this table.

• A reference video is cheapest per output second on long clips: the fixed 15 assumed seconds are spread over more output.
• At 5 seconds the reference input more than doubles the estimate, so use a reference image if a still carries the look you need.
• This is an estimate used at reserve time. The poll response usage.cost is the amount billed, so compare the two on your first few jobs.

Sending one

On /v1/videos, a reference video goes in input_references as a video_url item. Mode is inferred: references present means reference-to-video.

{
  "model": "seedance-2.5",
  "prompt": "Same camera move, new product on the table",
  "duration": 10,
  "resolution": "720p",
  "aspect_ratio": "16:9",
  "input_references": [
    { "type": "video_url", "video_url": { "url": "https://example.com/ref.mp4" } }
  ]
}