Announced 7 August 2026. Specifications below are from the platform's launch communication; pricing details are pending official publication. Verify against the Volcano Engine site before committing budget.
Seedance 2.5 is live. Three changes matter, and one of them is a bigger shift than the version number suggests.
What actually shipped
| Capability | What it does |
|---|---|
| Native 30-second output | Double the previous duration in a single generation, extendable across multiple rounds into several minutes of coherent video |
| 50 reference inputs | Up to 30 images, 10 videos and 10 audio tracks combined in one generation |
| Editable output | Generated video can be re-edited rather than only regenerated |
Seedance 2.5 is already listed by Perseworks, which sells Seedance generation by the second with no subscription commitment — useful if you want to test the new capabilities against your own footage before committing to a platform.
Why 30 seconds is a workflow change, not a spec bump
A 30-second spot is the standard unit of broadcast advertising. Until now, producing one meant generating shorter segments and stitching them, which is where continuity breaks: lighting drifts, a character's jacket changes weight, a camera move restarts instead of continuing.
Generating the full duration natively removes the seams entirely, because there are no joins to hide. Multi-round extension into minutes then covers the longer formats — brand films, training modules — where the alternative was assembling a dozen clips and accepting the drift.
Why 50 reference inputs matters more than it sounds
Reference stacking is how you hold a character, a product, or a brand look steady across a campaign. Thirty images plus ten videos plus ten audio tracks in a single generation is a large conditioning budget — enough to carry a face from several angles, a product from several sides, a lighting reference, a motion reference, and a voice, all at once.
Practically: the more references a model accepts, the less you rely on prompt wording to hold consistency, and prompt wording is the least reliable control surface in the entire pipeline.
The editability change is the significant one
This is the shift from generation to production.
Generative video has been a slot machine: you write a prompt, you get a take, and if one element is wrong — a gesture, a background detail, a beat of timing — your only recourse is to reroll the whole thing and hope the parts you liked survive. That is not how any professional post workflow operates. Editors work by adjusting what exists.
Second-pass editing means a generated clip becomes a workable asset rather than a finished artefact you accept or discard. For commercial work, where a client's note is usually "the same, but with one change", this is the difference between a tool you can quote a delivery date on and one you cannot.
Where it fits
- Advertising and TVC — a 30-second spot in a single native generation, then refined rather than rerolled.
- Brand content — reference stacking holds product and identity consistent across a series.
- Industrial and training material — multi-round extension covers long-form explanatory video, where continuity matters more than spectacle.
A pricing move is coming for Seedance 2.0 — but it is not live yet
Alongside the 2.5 launch, a promotion on the Seedance 2.0 line is being prepared:
| Tier | Indicated promotional pricing |
|---|---|
| 2.0 Fast | 75% of list price (a 25% reduction) |
| 2.0 Mini | As low as 40% of list price (a 60% reduction) |
This is not bookable yet. The promotion is still being staged, and both the start date and the final terms are subject to official publication by Volcano Engine. Treat the figures above as indicative, not as quotable rates. We will update this post when the official terms are published.
What this means if you are budgeting
The two announcements point in different directions, which is useful rather than confusing.
- 2.5 is for the hero shot — the deliverable where 30 seconds native, heavy reference conditioning, and the ability to fix one detail without rerolling are worth paying for.
- 2.0 Mini, if the promotion lands at the indicated level, becomes the drafting tier — storyboards, animatics, A/B tests, blocking. Work you throw away.
That split is the sane way to run generative video at scale: cheap iteration until the shot is decided, quality generation once. Our Seedance 2.0 pricing comparison sets out per-second costs across five platforms and is the place to start if you are working out what iteration actually costs you. Current per-second rates for both generations are listed at perseworks.ai/pricing.
Frequently asked questions
What is new in Seedance 2.5?
Three things: native 30-second generation extendable to several minutes across multiple rounds; up to 50 combined reference inputs (30 images, 10 videos, 10 audio); and the ability to re-edit generated video rather than only regenerate it.
How long can a Seedance 2.5 video be?
Thirty seconds natively in one generation, extendable through multiple rounds into several minutes of coherent output.
How many reference files can Seedance 2.5 take?
Fifty in a single generation — up to 30 images, 10 videos and 10 audio tracks.
Is the Seedance 2.0 discount available now?
No. It is in preparation. The indicated levels are 75% of list for Fast and as low as 40% of list for Mini, but the start date and final terms are subject to official publication by Volcano Engine.
Should I move from 2.0 to 2.5?
For finished deliverables, the editability and native duration are the reasons to move. For drafting and iteration, the cheaper 2.0 tiers remain the rational choice — particularly if the promotion lands as indicated.
Specifications are as communicated by the platform on 7 August 2026. Promotional pricing is indicative and unconfirmed; the Volcano Engine site is the authority on final terms and availability. Seedance 2.5 and 2.0 access is available through perseworks.ai. See also our per-second pricing comparison across five Seedance platforms.