Seedance 2.5 is out: what shipped, and where you can actually use it

What ByteDance shipped

Seedance 2.5 keeps 2.0's unified audio-video generation architecture and pushes on three fronts the Seed team names explicitly: long-form narrative, multimodal reference and editing. The framing in their own announcement is that users stopped wanting “a clip” and started wanting “a finished piece”.

The three capabilities, concretely

1. Thirty seconds, then minutes

A single generation now runs 30 secondsand organises several logically connected shots inside it — setup, development, turn, resolution — rather than extending one image. Multi-round extension then continues from a finished clip while holding the subject, setting and pacing, so several minutes of material share one audiovisual language. ByteDance's demo of this is a continuous take that follows a singer from her dressing room, through the backstage corridor, onto the stage and out to a full arena.

2. Fifty reference assets in one shot

30 images, 10 videos and 10 audio clipscan travel with a single request — enough to cast an entire scene. Their concert example references a hall, a pianist, a cellist, a violinist, a lead singer, six orchestra sections, four choir groups and four audience blocks, all in one prompt, and keeps each subject's appearance and voice stable in the same frame.

3. Editing addressed by timestamp

You specify what happens in which second at generation time, and after generation you can point at a segment and change the character, the action, the sound or the plot beat while continuity holds. Around that sit green-screen editing (swap the world behind a preserved subject, with cloth, hair, gait and light re-rendered for the new physics), viewpoint and camera-move editing (keep the performance, redesign the camera), and reference editing.

Where you can use it — re-checked September 18

• Volcano Ark API (official) — live. doubao-seedance-2-5 sits in the Ark model catalog. Our monitor first recorded it there on August 7, which is the real API opening date — a week after the model launched and three weeks after the date everyone reported.
• fal.ai— live, three public endpoints (text-, image- and reference-to-video). fal's own catalog quotes roughly $0.2205/s at 480p and $0.4730/s at 720p.
• kie.ai — live, with published credit rates of 17–158 credits per second depending on tier.
• BytePlus ModelArk (international) — still not carrying 2.5; the catalog stops at the 1.x and 2.0 families.
• Jimeng AI (即梦) and Doubao Pro (豆包专业版) — the consumer route, where 2.5 shipped first on launch day.

What it costs today, per channel

Published rates, re-checked daily by our monitor. The spread above is not a historical footnote — it is the current price list:

What it looks like

Two clips ByteDance published with the launch. The first is the headline capability in one file — thirty seconds, one take, no cuts, with sound. The second is the short film the Seed team says Seedance 2.5 generated on its own, and it is the more revealing of the two: four and a half minutes of dialogue-driven scenes built by extension, which is exactly the workflow 2.5 was designed for.

Official Seedance 2.5 demos · ByteDance Seed · source · mirrored and re-encoded here, sound intact.

Official prompts from the launch

ByteDance published its demo prompts, which tells you more about how 2.5 wants to be driven than any spec sheet. Two patterns stand out.

Timestamps carry the direction. The Peking-opera demo is written as a schedule — “0–5s: opening close-up, camera orbits and transitions to a mid shot… 6–10s: stable orbit around the water sleeves… 11–20s: the martial role enters with a backflip” — with each beat naming which reference image supplies which character.

References are addressed like variables. Prompts point at @图片 1 (image 1), @视频 1 (video 1) and @白模 1 (white model 1) by number, then say what each one governs — this one supplies the camera path and blocking, that one supplies material, lighting and colour. The full set of launch prompts, translated and annotated, is in our Seedance 2.5 prompt library.

Where it is being used already

Two industry threads came with the launch. In education, Doubao Classroom is already generating lesson video — turning historical context, scientific principles and experiment procedures into moving demonstrations. In industry, the pitch is synthetic training data: robotics perception, industrial simulation and autonomous-driving scenarios, including rare events like extreme weather and unusual road conditions that are hard to collect in the real world.

Seedance 2.5 or MiniMax H3?

Both launched on July 31, and the contrast is unusually clean:

Longer and more referenceable versus cheaper by the second. Full breakdown on the MiniMax H3 tracker, and all three together in Seedance 2.5 vs H3 vs FLUX 3.

Get the price moves, not the noise

One email when a channel adds Seedance 2.5 or changes its rate — fal doubled MiniMax H3's price on launch day and later cut it back, and that is the kind of move that decides which endpoint you call. We check the Ark and BytePlus catalogs, fal and kie every day.