Seedance 2.5 reference video: dearer at 10 s, cheaper at 30 s

Adding a reference video to a Seedance 2.5 request raises the price of short clips and lowers it for clips longer than about 23 seconds. Sume's estimate adds an assumed 15 seconds of input video to the output length, then multiplies the total by 0.6, so the reference-video price beats the no-reference price once the output passes 22.5 seconds.

This is the 2.5 version of a rule that an earlier post covers for Seedance 2.0. It matters more on 2.5 because the model runs to 30 seconds, which is where the crossover shows up. The figures below come from Sume's pricing code and the Video generation docs; fal's own page for the model was read on 2026-10-02 to confirm the rate and the 0.6 multiplier.

How does Sume price a Seedance 2.5 clip with a reference video?

Seedance is billed in video tokens. fal's reference-to-video page gives the formula as output height x output width x (input video seconds + output seconds) x 24 / 1024, at $0.0214 per 1,000 tokens for 480p and 720p, with a 0.6 multiplier when video references are present. Sume's pricing code follows the same shape, uses $0.0234 per 1,000 tokens at 1080p, and bills the provider list times 1.25.

The difference is the input length. Sume does not probe your reference clip before submitting, so it assumes 15 input seconds whenever any reference video is present, whether the file is 3 seconds or 14. Reference images and reference audio add nothing to the estimate; only a video reference triggers the extra seconds and the 0.6 rate.

• No reference video: price = output seconds x per-second rate.
• Reference video present: price = (output seconds + 15) x per-second rate x 0.6.
• Break-even: (d + 15) x 0.6 = d gives d = 22.5 seconds.

What does the crossover look like in dollars?

Each figure is the Sume estimate for a 16:9 clip, list times 1.25 rounded up to the cent. The 720p column is the one most requests use.

Should you add a reference video to save money?

No. The saving is a side effect of an estimate, not a pricing tier, and it exists only if you already want a clip of 23 seconds or more. Choose the reference video for what it does: it carries motion or camera work into the output, as described in the Seedance 2.5 camera-move post.

The sharper practical point is the short end. A 5-second Seedance 2.5 clip at 720p estimates at $2.89 alone and $6.94 with a reference video, about 2.4 times as much, because 15 assumed seconds are being added to 5 real ones. If you only need a style cue, a reference image gets it without that jump.

Is the estimate the same as what gets charged?

Sume reserves the estimate when the job is admitted and settles on the same admit, reserve and capture path as other generation jobs, per the Video Router docs. The catalog's billable_formula field on each model row is the per-model truth for how list becomes billable, so read it from GET /v1/video-router/models/seedance-2.5 rather than trusting a blog table when rates move.

Two caveats. fal's page counts the real input-video duration, while Sume assumes 15 seconds, so a 4-second reference clip is priced by Sume as if it were 15. And the table assumes 16:9; other ratios use different pixel dimensions, which changes the token count.

How do you check the price before you submit?

Read the catalog row for the model, take list_usd_micros_per_1000_video_tokens, and apply the formula above with your resolution's pixel size. Sume's own numbers for 720p 16:9 are 1280 x 720 pixels, which is 21,600 tokens per output second. Multiply by the rate and by 1.25 to reproduce the 720p column.

For a longer film, see the multi-round extension post, which prices 60 and 90 seconds as several 30-second rounds. ByteDance's launch post describes multi-round extension as the route beyond 30 seconds, and it lists the 30-image, 10-video, 10-audio reference ceiling that Sume does not expose in full (see the reference limits post).