What Seedance 2.5 is
Seedance is a line of video models by ByteDance, the company behind TikTok and CapCut. Version 2.5 was announced on 23 June and released on 31 July 2026. The architecture is the same as in Seedance 2.0: the model generates video and sound together rather than laying audio on top.
The developer frames the update like this: the model should produce a finished piece, not an isolated clip. Hence three big changes.
What is new compared to 2.0
| Seedance 2.0 | Seedance 2.5 | |
|---|---|---|
| Length in one pass | up to 15 seconds | up to 30 seconds |
| References | around 12 | up to 30 images, 10 videos, 10 audio |
| Fixes | full regeneration | targeted editing by timecode |
| Extending a clip | no | several rounds of extension |
| Sound | native | native |
30 seconds in one piece
In 30 seconds the model has room to show a short story rather than stretch a single moment: a setup, a development and an ending. In the official demo a singer walks from the dressing room through a corridor onto the stage in one continuous shot. Previously such a clip was assembled from three segments with all the pain of matching the seams.
Extension was added too: you can append the next 30 seconds to a finished clip with the same character and setting. That is how multi-minute videos get built without manual stitching.
Many references
Up to 30 images, 10 videos and 10 audio files in a single request. What this means in practice: you can define the appearance of several characters, the location, the style and even a voice, and the model will hold all of it in frame. For scenes with several characters this is the main breakthrough, because character consistency used to break down precisely on group shots.
Editing instead of regenerating
You can point at a timecode and ask to change only that fragment: what the character does, how the camera moves, the background. Everything else stays as it was. Previously one bad second meant regenerating the whole clip and paying for it again.
Separately, the developer claims more natural skin, eyes and lighting, meaning less of the plastic gloss that gives AI video away in the first frame.
An important caveat. The capabilities above are the ones stated by the developer. The feature set on a particular platform may differ: extension, video references and timecode editing are not wired up everywhere yet.
What Seedance 2.5 costs on KADR
Prices in tokens, sound does not affect the price, October 2026:
| Resolution | 6 seconds | 30 seconds |
|---|---|---|
| 480p | 102 (~161 rub.) | 502 (~794 rub.) |
| 720p | 216 (~341 rub.) | 1075 (~1700 rub.) |
| 1080p | 537 (~849 rub.) | 2683 (~4242 rub.) |
This is not a public offer.
For comparison, Seedance 2.0 on KADR: 6 seconds at 480p costs 69 tokens, at 1080p 362 tokens. So 2.5 is roughly one and a half times more expensive. A minute at 1080p on Seedance 2.5 works out at about 8500 rubles, one of the most expensive options in the catalogue. Comparison with the other models is in the article on how much a minute of AI video costs.
Relaxed moderation filter
Seedance is known for strict moderation: the model often refuses if the request involves a realistic face, even when you simply want to animate an ordinary photo. On KADR, Seedance 2.5 runs with a relaxed filter: it is always on, and the prices in the table already account for it (roughly 10 percent above the base rate).
What that gives you in practice: fewer false refusals on ordinary requests involving people. What it does not give you: the right to generate real people without consent, celebrities or someone else’s characters. Such requests can still be rejected, and rightly so.
When 2.5 pays off and when it does not
Take Seedance 2.5 if:
- you need a coherent clip longer than 15 seconds with no seams (narrative ads, a music clip, a scene that develops);
- there are several characters in frame and each must stay themselves;
- you are fixing a finished clip and do not want to pay for a full regeneration;
- natural skin and lighting matter in close-ups.
Seedance 2.0 or cheaper is enough if:
- the clip is assembled in an editor from short 5 to 8 second scenes, like most reels;
- you are still tuning the prompt and running drafts;
- there is one character in frame doing something simple.
The money-saving routine is the same as with any expensive model: drafts on a cheap model at 480p, the final take on Seedance 2.5. If you are unsure which model fits your task, there is a separate breakdown on which AI video model to choose.
How to write a prompt for 30 seconds
A long clip needs a different prompt. A single line like «a girl walks down the street» will turn into half a minute of pointless walking.
Write by timecode, the way the developer does in their own examples:
- 0 to 8 seconds: where we are, who is in frame, how the camera sits.
- 8 to 20 seconds: what happens, how the camera moves.
- 20 to 30 seconds: how it ends, the final shot.
One action and one camera move per segment. General prompt rules are covered in the article on how to write prompts for video.
Ready prompts by timecode: Seedance 2.5 prompts. A comparison with Kling and Veo over 30 seconds: Seedance 2.5, Kling or Veo.
In short
Seedance 2.5 is a noticeable step forward: 30 seconds in one piece, up to 50 references and edits to individual fragments. You pay roughly one and a half times more than for 2.0. Take 2.5 for narrative clips, group scenes and final versions, and do drafts and short scenes on cheaper models.
