XinYu AI Testing Team · Published Sep 30, 2026 · updated Nov 1, 2026
Before you generate, the panel shows how many credits the clip will cost, and that number is computed by the same function that does the actual charge. But one total doesn't tell you where it comes from, so it doesn't tell you which setting to change to spend less.
This post breaks Seedance 2.5 pricing down one setting at a time. Every item is given as a relative change, not a price. Prices move when the upstream changes them; the ratios mostly don't. Every ratio here was computed by calling the same pricing functions our billing uses in production. None of them are estimates.
The whole thing in one table
| What you change | What happens to the price |
|---|---|
| Duration | Linear: 10s ≈ 2× a 5s clip, 30s ≈ 6× |
| Resolution | 480p ≈ 0.45× 720p; 1080p ≈ 2.46× 720p |
| Audio on | No change |
| Reference images (one or several) | No change |
| Reference audio | No change |
| A short reference video | Almost no change (+0% to +8%) |
| A reference video longer than ~2/3 of the output | Each extra reference second costs about 0.6 of an output second |
| Video edit | Priced on the source length, about 1.2× a text-to-video clip of the same length |
| Video extend | Priced on the length you pick plus the source length, same rules as a reference video |
Duration: linear per second
Seedance 2.5 bills by token, and tokens = pixels per frame × frames, at a fixed 24 frames per second. So price tracks duration almost exactly:
| Duration | 4s | 5s | 10s | 15s | 20s | 30s |
|---|---|---|---|---|---|---|
| Relative to 5s | 0.80 | 1.00 | 1.99 | 2.99 | 3.99 | 5.98 |
Short clips cost a hair more per second because every clip carries one extra first frame (frames = 24 × seconds + 1). On a short clip that frame is a slightly bigger share, but the difference is under 1%. You can pick any whole number of seconds from 4 to 30.
Resolution: 1080p is 2.46×, not 2.25×
1080p has 2.25× the pixels of 720p, but 1080p also has a higher price per million tokens. The two multiply to 2.46×. 480p has about 0.44× the pixels of 720p at the same per-token price, so it comes to about 0.45×.
While you're still testing a prompt or a camera move, 480p costs less than half of 720p. One caveat: regenerating at a different resolution won't give you the same shot. 480p is for checking whether a prompt can get you what you want. It isn't a preview of the final frame.
Aspect ratio: for text-to-video and reference generation, aspect ratio doesn't change the price. Edit and extend take the output aspect ratio from the source clip, and pixel counts differ slightly between ratios: under 1% at 720p, up to about 5% at 480p.
Audio, reference images and reference audio cost nothing extra
- Audio: a clip with sound costs the same as one without.
- Reference images: one image and five images cost the same. Images aren't part of the price formula.
- Reference audio: also left out of the price. It only has limits on count and total length.
Reference video: short ones are nearly free, long ones add per second
Once you attach a reference video, billing switches to a different rule. The price per million tokens drops by about 40%, but the billed length becomes "output seconds + reference seconds", with a floor of 5/3 of the output length (rounded up).
Here's what that works out to (720p; each number is the price with the reference video ÷ the price without it):
| Output \ Reference | 2s | 4s | 6s | 8s | 10s | 15s | 20s | 30s |
|---|---|---|---|---|---|---|---|---|
| 5s | 1.08 | 1.08 | 1.32 | 1.55 | 1.79 | 2.39 | 2.98 | 4.17 |
| 10s | 1.02 | 1.02 | 1.02 | 1.08 | 1.20 | 1.49 | 1.80 | 2.39 |
| 15s | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.20 | 1.39 | 1.79 |
| 30s | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.20 |
How to read it:
- While the reference is no longer than about 2/3 of the output, you're billed at the floor, and the price stays close to no reference at all (+0% to +8%, most visible on 5-second clips). We worked out why the "40% cheaper" rate never actually shows up, and why it cancels out almost exactly, in Seedance's cheaper input-video tier saves you nothing.
- Past that line, each extra second of reference video costs about 0.6 of an output second.
- Reference length is rounded down, so a 4.9-second clip is billed as 4 seconds.
- With several reference videos, their total length counts. The total can be up to 30 seconds.
So keep only the part of the reference you need the model to see. A 15-second output with a 30-second reference costs 1.79×. Trim the reference to 10 seconds and you're back at 1.00×.
Video edit: priced on the source length, about 1.2×
An edit's output follows the source length (BytePlus's docs say the output is never longer than the source and at most 0.4 seconds shorter), so it's priced on the source length. The source also counts as input, so the billed length is twice the source. Multiply that by the roughly 40% cheaper rate and you get about 1.2× a text-to-video clip of the same length. The source must be 4 to 30 seconds long.
Keep in mind that an edit's price depends only on how long the source is, not on how much of it you change. Removing one cup from the start of a 30-second clip is still billed as 30 seconds.
Video extend: the length you pick, plus the source
With extend, you choose the output length (4 to 30 seconds), and the source counts as input under the same rules as a reference video. For a 10-second extension: a 5-second source is 1.02×, a 10-second source 1.20×, a 20-second source 1.80×.
Why can a Seedance 2.5 edit come out up to 0.4 seconds shorter than the input video?
Because that's how the model is specified. In edit mode the output duration tracks the input video, but it can come out slightly shorter, by at most 0.4 seconds. The 0.4-second figure is BytePlus's official wording, not a limit we measured. From the official docs:
Due to input frame processing, the output may be slightly shorter than the input video, with a difference of no more than 0.4 seconds.
"Input frame processing" is the only reason the docs give. The same docs also say the output may not be a whole number of seconds. (Quoted from the BytePlus docs as of November 1, 2026. The multipliers in this section were computed with our production pricing functions on the same day.)
We ran two edits ourselves (n=2) and worked out the output length from the upstream usage. The edit on a 5.04-second source came out the same length as the source. The one on a 4.5-second source came out 3 frames short, about 0.125 seconds. Both are inside 0.4 seconds. Two runs can't tell you how often an edit comes out short, or by how much.
You can't set the length of an edit, either. For edit tasks the official docs only allow duration to be -1, so length and aspect ratio both follow the source. If you need a longer clip, use extend.
So if you're lining an edit up against the original or a music track, go by the output's actual length. Don't assume it matches the source. For how to write the edit prompt itself, see Writing Seedance 2.5 video edit prompts: removing is easy, replacing is hard.
Edit vs extend: who sets the length, and how many seconds are billed
| Video edit | Video extend | |
|---|---|---|
| Output length | Follows the source, up to 0.4s shorter (official) | You pick, 4–30s |
| Output aspect ratio | Follows the source | Follows the source |
| Billed length | Source + source, about 2× the source | Your output length + source length, at least 5/3 of the output |
| Rate | The video-input rate, about 40% cheaper | Same |
| vs. text-to-video of the same length | About 1.2× | About 1.0–1.1× while the source is under ~2/3 of the output, then more per second |
| Source length (official) | 4–30s | 2–30s per video, 30s total |
A few notes:
- An edit is billed on the source length, not on however long the output turns out.
- The copy of the source that counts as input is rounded down to whole seconds. A 4.5-second source is billed as 4 + 4.5 = 8.5 seconds, not 9, so an edit on a source with a fractional length comes in slightly under 1.2×.
- Extend uses exactly the same rule as a reference video, so you can read it off the reference video table above: the column is your source length, the row is the output length you picked.
- The "2–30s per video, 30s total" for extend is the official general rule for tasks other than editing. The docs don't list a separate limit for extend.
- You don't need to enter the source length. An edit whose source isn't 4–30 seconds, or a request whose source videos add up to more than 30 seconds, is stopped before you're charged.
Edit and extend: common questions
Can I set the output length in Seedance 2.5 edit mode?
No. The official docs say duration only supports -1 for edit tasks, so you can't specify a length. Length and aspect ratio follow the source. To change the length, use extend, where you pick anything from 4 to 30 seconds.
Can I edit a clip shorter than 4 seconds or longer than 30?
No. The official docs require the video being edited to be 4–30 seconds long and return an error otherwise. On XinYu AI, an edit outside that range is rejected before you're charged. For tasks other than editing, the official limit is 2–30 seconds per reference video.
Does extending by 10 seconds cost the same as generating 10 seconds from a reference video?
Almost. Both bill output seconds plus source seconds (rounded down), with the same floor of 5/3 of the output. With the same source, output length and resolution, the pricing formulas differ only in aspect ratio: extend is priced on the source's aspect ratio and a reference generation isn't. Pixel counts differ slightly between ratios, under 1% at 720p and up to about 5% at 480p.
If a generation fails
A failed generation refunds its credits. For what happens when a clip is rejected by moderation, see When a generation is rejected, do you get your credits back?
Where the numbers come from
Every ratio in this post was computed by calling the pricing functions our production billing uses. The price shown in the panel and the amount actually charged go through the same function. Figures are as of September 30, 2026.
