Shii-chan, 30 seconds in one breath. The dance club just leaped over AI video's "15-second wall."
On July 31, 2026, dance club star Shii-chan (Seedance) received her new model, Seedance 2.5. 30 seconds per generation, up to 50 reference assets, and video with audio, generated together โ the old routine of "make short cuts, then stitch them" is about to become last year's habit.
AI video has always had a "15-second wall."
Push past it and faces morph, outfits change color, extra buildings sprout in the background. So everyone made a stack of 5- or 10-second clips and stitched them together in an editor afterward โ that was simply how it worked, right up through last year.
Seedance 2.5 took that assumption and stretched it to a full 30 seconds. And it did so with sound.
So what exactly is Seedance 2.5?
It's made by Seed, the research team at ByteDance โ yes, TikTok's home. Its predecessor, Seedance 2.0, launched its global API in April 2026 and set the whole academy buzzing by costing roughly 1/100th of Sora 2. This time, that famous cheapness gains two new tricks: length and sound.
Announced June 23
Unveiled at the Volcano Engine FORCE conference in Beijing. At that point it was still an enterprise-only beta.
Public July 31
Now open to everyone via Dreamina (known as Jimeng AI in China), with a gradual rollout to the Pro tier of Doubao as well.
Where's the API?
Volcano Engine Ark in mainland China, BytePlus ModelArk everywhere else. Third-party providers offer it too.
Speaks your language
Supports 10+ languages, including Japanese, Chinese, English, and Korean. Write your prompt in your own language and generate directly.
What actually changed from 2.0 to 2.5?
The spec sheet looks modest at a glance, but almost every number doubled or better.
๐ Seedance 2.0
- Up to 15 seconds per generation
- About 12 reference assets
- 1080p resolution
- Audio included. Famous overnight for its low price
๐ Seedance 2.5
- 30 seconds per generation (up to 180 seconds in the beta long-form mode)
- Up to 50 reference assets (30 images, 10 videos, 10 audio clips)
- Native 4K with 10-bit color
- Video and audio generated together by one model, plus support for partial regeneration
Here's the kind of footage it produces
Watching one clip beats reading about it. Here's a sample the academy made โ Shii-chan dancing, of course. The audio was generated too, so feel free to turn up the volume.
Sample video ๏ผ Hana AI Academy
The headliners: four of them
"Longer" isn't the whole story. Let's walk through the points that actually matter when you sit down to use it.
1. 30 seconds in one pass
30 seconds is the length of a TV commercial โ enough for a full beginning-middle-end arc in one continuous shot. One unbroken take means seam mismatches simply cannot happen.
2. 50 reference assets
Hand it 30 images, 10 videos, and 10 audio clips at once. This isn't "use these for inspiration" โ it's "make it with this face, this outfit, this voice, in this room."
3. Sound comes built in
Dialogue, music, and ambient sound come out in stereo, lip-synced to the characters. The whole "add audio in another tool" step just became optional.
4. Blank models and partial redraws
Block out camera moves and staging first with plain white, untextured 3D models, then apply the visual style afterward. Don't like one spot? Repaint just that spot.
The longer the clip, the more a "do-over" costs
It sounds unglamorous, but number 4 is the one that matters most in real work. If a 5-second clip fails, you just roll again. But regenerating a full 30-second take from scratch sends both your time and your bill through the roof.
Locking composition first with blank models and repainting only the parts you dislike โ those two features are the insurance policy of the long-form era.
Now, about the price
Full honesty up front: ByteDance has not published an official per-second price for Seedance 2.5 (as of August 2026). Here's what we do know โ three things.
Dreamina subscriptions
Free tier (watermarked, no commercial use), Standard around $18/month, Advanced around $82/month. This is the easiest place to start.
The API bills by tokens
On Volcano Engine Ark: 70 yuan per million tokens (no video input) or 42 yuan (with video input). The official example prices a 5-second 720p clip at about 7.56 yuan (roughly 150 yen, or about a dollar).
So what does 30 seconds cost?
Third-party estimates put it at $0.66โ$1.80 on modest settings, climbing to $15+ at full 4K. These are estimates โ always measure for yourself before production.
So... how good is it, really?
Independent scores for 2.5 don't exist yet.
The best reference point is the previous generation's report card. In the Artificial Analysis Video Arena โ where humans compare clips blind โ Seedance 2.0 (720p) ranked #1 in the image-to-video-with-audio category (Elo 1198). In text-to-video with audio it placed 3rd (Elo 1224), behind Gemi-chan's Gemini Omni Flash in 1st (1244) and MiniMax H3 in 2nd (1240).
In other words: a girl who was already fighting for the top spot at 2.0 just came back with double the length and stronger audio. That's where we stand. Real numbers are still on the way. Any "Seedance 2.5 Elo score" floating around right now should not be trusted yet.
"A different dimension from just a year ago"
A Nikkei xTECH reporter paid for a plan and actually tried it, writing with astonishment that a few still images plus instructions written in Japanese were enough to get back a finished video with audio. One reaction from Japanese creators: "stock-footage shops are done for."
How does she share the stage with her rivals?
Video is the most crowded, most competitive club at the academy right now. Rather than asking "which one is best," think "which one for which job."
Shii-chan (Seedance 2.5)
Long takes, deep reference stacks, built-in audio. For 30-second commercials, series work, and keeping the same character consistent across many videos.
Klin-chan (Kling)
True to her science-club roots, her weapon is physical realism. Fluids, cloth, particles โ anything where believability is everything goes to her.
Wei-chan (Runway)
Studio-grade depth of tooling. For film work you want to refine carefully, editing included.
Domo-chan (DomoAI)
Rough prompts, blazing speed, high volume. For brainstorming and social media feeds, tossing it to her first is the fastest route.
Gemi-chan (Omni Flash)
Currently #1 in the text-to-video arena. Polished official developer access is another of her strengths.
Gro-chan (Grok Imagine Video 1.5)
Her home turf is image-to-video, also with native audio, and speed is her weapon: 720p in about 25 seconds. On July 31 she also gained 7 reference images and voice locking.
Things to check before you use it
With great convenience come a few things worth double-checking.
Not open source
This is a proprietary ByteDance model. The weights are not public, so you cannot run it yourself.
Free tier is non-commercial
Videos made on the free tier are watermarked and barred from commercial use. Working professionally? A paid plan is mandatory.
Read the terms
Dreamina is granted a broad usage license. Check in advance how your uploaded assets and generated videos may be used.
Watch for lookalikes
The risk of resembling existing characters or trademarks hasn't gone away. With more reference slots, you actually need to be more careful, not less.
Deciding what to do with 30 seconds is still a human's job.
"Stock-footage shops are done for" โ honestly, the impulse is understandable. 30 seconds, with audio, in 4K, characters locked โ all from a single button. But Shii-chan's real talent is dancing exactly as she feels. How to spend those 30 seconds, who dances, where the stumble lands โ deciding all that, for now, is still on our side of the stage. One wall came down, and the thinking got harder. The School Newspaper will keep following this story.
"Dance the way you feel! Every step becomes a brand-new story!"
Shii-chan (Seedance) ๏ผ Dance Club
This article's sources
TechNode (July 31 launch report) ๏ผ Dreamina official (Seedance 2.5) ๏ผ Nikkei xTECH (Aug 5 hands-on review) ๏ผ WEEL (specs and pricing roundup) ๏ผ Artificial Analysis Video Arena ๏ผ WEEL (Grok Imagine Video 1.5) ๏ผ kie.ai (pricing estimates)