Seedance 2.5 is ByteDance's longest-take video model: up to 30 seconds in one pass, up to 50 image, video and audio references, generated sound and localized editing. It ranks 1st in image to video and 5th in text to video on Artificial Analysis, but tops out at 720p on public APIs and costs roughly $0.10 to $0.47 a second depending on where you run it. The 4K version Volcano Engine launched on October 23, 2026 is trial-only for selected enterprise customers.
This Seedance 2.5 review answers one question: is ByteDance’s newest video model worth using, and what does a clip really cost? We read ByteDance’s page, three API providers, Artificial Analysis and the press coverage of the October 4K launch, with prices checked November 6, 2026.
We didn’t run our own generations for this review. Every quality claim below comes from ByteDance’s own material or a named third party, and we say which.
It’s for creators, ad teams and developers choosing a video model. If you want rivals instead, see our Seedance alternatives list.
TL;DR: Seedance 2.5 review verdict
Seedance 2.5 is the strongest long-take video model available right now. It makes 4- to 30-second clips with sound, takes up to 50 references, and leads Artificial Analysis in image to video. The costs are real: 720p on public APIs, and a price that runs from about $0.10 to $0.47 a second.
| Category | Verdict | Why |
|---|---|---|
| Clip length | Best in class | 30 seconds in one pass, double Seedance 2.0 |
| References | Strong | Up to 50 files: 30 images, 10 videos, 10 audio clips |
| Leaderboards | Strong | 1st in image to video, 5th in text to video on Artificial Analysis |
| Resolution | Limited | 480p and 720p on public APIs; 4K is enterprise-trial only |
| Price | Mixed | $0.10 to $0.47 a second at API rates |
| Independent testing | Thin | ByteDance publishes demos; neutral long-form tests are scarce |
Use it for takes over 15 seconds, reference-heavy ads and branded clips. Skip it if you need 1080p today or the cheapest possible 720p: Seedance 2.0 is cheaper per second.
What is Seedance 2.5?
ByteDance Seed’s Seedance 2.5 page, checked November 6, 2026
Seedance 2.5 is ByteDance’s follow-up to Seedance 2.0, which arrived in early 2026. ByteDance Seed calls it “a next-generation audio-video joint generation model, built for 30-second storytelling with precise reference control and powerful editing capabilities.”
Launch dates differ by source. CineD and several outlets place the announcement at Volcano Engine’s FORCE conference on June 23, 2026. Artificial Analysis lists the model as released in July 2026, and OpenRouter shows August 7, 2026.
ByteDance sells it through its own products (Dreamina, Jimeng, CapCut) and through BytePlus ModelArk for developers. Third parties such as fal and OpenRouter resell it.
Seedance 2.5 specs at a glance
These are the limits third-party API providers publish. ByteDance’s own pages don’t list a full spec sheet, so the table cites the source of each row.
| Spec | Seedance 2.5 | Source |
|---|---|---|
| Clip length | 4 to 30 seconds, whole seconds | fal, OpenRouter |
| Resolution | 480p or 720p | fal, OpenRouter |
| Aspect ratios | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 | OpenRouter |
| Audio | Generated with the picture, lip-synced dialogue | fal |
| References | Up to 50: 30 images, 10 videos, 10 audio clips | fal |
| Reference length | Video and audio 1.8 to 30.2 seconds | fal |
| Modes | Text, image, first/last frame, reference, edit, extend | fal, OpenRouter |
| Speed tier | No Fast version (Seedance 2.0 has one) | fal |
On fal, turning audio off doesn’t lower the price. Editing and extending both run through the reference-to-video endpoint, with no separate model.
Seedance 2.5 key features
30 seconds in one pass
This is the headline. ByteDance says you can “create videos up to 30 seconds in a single generation”, and its page mentions extending a clip twice beyond that.
Most rivals stop at 10 to 15 seconds. A 30-second take means one generation where you’d otherwise stitch two or three clips, and the stitching is where character drift and lighting jumps creep in.
Up to 50 references
Seedance 2.0 took up to 12 references. Version 2.5 takes up to 50 across images, video and audio, according to fal’s comparison.
ByteDance says it reads reference footage for “intent, framing and cinematic language”, not just motion. You can pull identity from an image, camera work from a clip and voice from audio in one request. Image and audio references are free on fal; reference video is billed by its length.
Editing and camera control
ByteDance’s page lists three control tools: white-model control, green-screen editing and “professional camera movement and performance blocking”. CineD describes the white-model feature as a 3D blockout that sets camera position and staging, and calls it the first of its kind in a video model.
It also describes localized redraws that change part of a frame while keeping performance, lighting and camera behavior. fal and Runware document the same edit mode on their endpoints, including removing or replacing elements and swapping backgrounds.
Sound and lip-sync
Seedance 2.5 generates audio in the same pass as the video. Dialogue you put in quotes comes back lip-synced, per fal, and Runware lists support for more than 10 languages with language switching inside one clip.
Multi-shot and character consistency
ByteDance doesn’t publish a multi-shot spec for 2.5, so treat this as claimed rather than proven. Dreamina says the model can “keep the same character identity, face, outfit, body language and movement logic across multiple AI video shots” when your references are organized.
fal tested a four-cut, 10-second night market sequence on both Seedance versions, but didn’t say which handled the cuts better. CineD’s advice is blunt: check character identity and lighting across the full 30 seconds yourself, because launch reels are curated.
Character swap
ByteDance’s own page doesn’t name a person-swap tool. What it does list is reference-driven editing, and the API docs describe removing or replacing elements in an existing clip.
Dreamina’s “swap characters” landing page reads as marketing: its steps describe picking Seedance 2.5 and uploading an image, which is generation rather than footage editing. Its game-animation page says localized editing can swap or remove armor and other details on a generated character. If you need a clean person swap in real footage, test it on a short clip before you commit.
In Vidwip, Seedance 2.0 Fast, 2.0 and 2.5 can swap a character into a clip you bring, as the models page notes, and the dedicated AI UGC tool runs Kling O3 Edit, MiniMax H3 and Runway Aleph 2.0 for the same job.
The 4K Volcano Engine variant
Volcano Engine, ByteDance’s cloud arm, launched Seedance 2.5 4K on October 23, 2026, aimed at film and TV production. TechNode, citing IT Home, reports extra training on “facial expressions, subtle movements and visual details”.
It supports 10-bit color and a draft mode: you test shots and refine storyboards at cheaper 480p, then render the final in 4K. It’s also meant to hold image quality and color when you work from references or edit footage.
Access is the catch. TechNode says it is “currently available for trials by selected enterprise customers”, with unnamed film and post-production companies in China and overseas testing it.
Pricing isn’t published, and public APIs still stop at 720p. CineD noted that ByteDance’s native 4K, 10-bit output was confirmed for Seedance 2.0, and fal lists 2.0 at up to 4K. So if you need 4K today, Seedance 2.0 is the model you can actually call.
Seedance 2.5 pricing: cost per clip
OpenRouter’s Seedance 2.5 listing, checked November 6, 2026
Seedance 2.5 is billed per second of output, and the rate varies about fourfold between providers. All figures are US list prices checked November 6, 2026, for 16:9 clips with audio and no video reference.
| Where | 480p | 720p | 30 s at 720p |
|---|---|---|---|
| BytePlus ModelArk (via CometAPI’s quote) | $0.103/s | $0.231/s | $6.93 |
| OpenRouter | from $0.1028/s | not itemized | not itemized |
| fal | about $0.2205/s | about $0.473/s | about $13.87 |
| Vidwip | not offered | $0.28–$0.35/s | 1,050 credits, $8.30–$10.50 |
| Dreamina (consumer app) | not itemized | 37 credits/s | 1,110 credits |
BytePlus bills by token, at $10.70 per million tokens without video input and $6.40 with video input, as CometAPI’s pricing guide reports. CometAPI’s per-second figures are its own conversion, and it adds that BytePlus charges only for successfully generated videos.
fal bills $0.0214 per 1,000 tokens, and a reference video drops the 720p rate to about $0.2838 per billed second. fal calls its per-second figures a ceiling.
In Vidwip’s AI video generator, Seedance 2.5 is 35 credits a second at 720p with sound, for 4 to 30 seconds. A credit costs $0.0079 on the Pro annual plan and $0.01 on Starter monthly, so a second runs $0.28 to $0.35 and a 10-second take is 350 credits. See Vidwip pricing for plans.
Dreamina’s consumer app meters Seedance 2.5 at 37 credits a second at 720p, against 17 for Seedance 2.0. Its US Basic plan advertises a $0.29 first month, then $15 a month, and the page notes the plan is credit-capped.
Cost per accepted clip matters more than cost per second. CometAPI’s example: if only 70 of 100 720p clips pass review, the effective price is about $1.65 a 5-second clip, not $1.16. Draft at 480p, lock the prompt, then render at 720p.
How Seedance 2.5 ranks
We checked Artificial Analysis’ Video Arena on November 6, 2026. It ranks models by blind human votes, which makes it the closest thing to a neutral test.
| Board | Seedance 2.5 | Seedance 2.0 | Leader |
|---|---|---|---|
| Image to video | 1st, Elo 1159 | 3rd, Elo 1135 | Seedance 2.5 |
| Text to video | 5th, Elo 1155 | 9th, Elo 1136 | Wan 3.0, Elo 1172 |
In text to video it sits just behind Wan 3.0 (1172), Utopai X (1169), MiniMax H3 at 768p (1160) and FLUX 3 Max (1156). The 95% interval on its 1155 is 1141 to 1169, so the top five are close to tied.
The takeaway: Seedance 2.5 is best when you start from an image, and one of five near-equals when you start from text. It’s also the only top-tier model here that does 30 seconds in a single pass. For more on the field, see our Hailuo 3 review and the list of open-source AI video models.
Seedance 2.5 pros
- 30-second single takes: double Seedance 2.0 and most rivals
- 50 references: images, video and audio together, with identity and camera work from separate inputs
- Top of the leaderboard in image to video: 1st on Artificial Analysis
- Sound built in: lip-synced dialogue, with audio references free on fal
- Editing tools: localized redraws, green-screen editing and a 3D white-model blockout
- Wide access: Dreamina, CapCut, BytePlus, fal, OpenRouter and Vidwip
Seedance 2.5 cons
- 720p ceiling on public APIs: the 4K variant is enterprise-trial only
- Pricey per second: $0.231 to $0.473 a second at 720p, more than Seedance 2.0
- No Fast tier: Seedance 2.0 has one for cheap drafts
- Claims without proof: ByteDance’s 20 percent prompt-adherence gain and multi-shot consistency have no independent test
- IP risk: cease-and-desist letters tied to Seedance 2.0 remain outstanding, per CineD, so use your own or licensed material
- Mixed launch story: announced in June, but sources disagree on when it reached the public
Who should use Seedance 2.5?
Seedance 2.5 fits ad teams, social creators and developers who need long, controlled takes with sound. Pick it for product films, music-driven clips and anything where one continuous 20- to 30-second shot beats stitched cuts.
Pick something else when:
- You need 1080p or 4K now: use Seedance 2.0
- You’re drafting a lot: use Seedance 2.0 Fast or a cheaper model
- You want an open model: see our open-source video model roundup
- You want a head-to-head: read Kling 4.0 vs Seedance 2.5 or Seedance 2.5 vs Veo 3.1
Seedance 2.5 in Vidwip’s model list: 4 to 30 seconds, 720p, with sound
If you want to compare Seedance 2.5 against other models on the same prompt, Vidwip’s AI video generator runs it beside Kling 3.0 Pro, Veo 3.1, Hailuo 3 and Wan 3.0. The models page lists all 16 video models with their lengths and resolutions. For the how-to, see how to use Seedance 2.5, and for rates by platform see the cheapest Seedance 2.5 provider.
Conclusion
Seedance 2.5 is the model to try when a clip has to run past 15 seconds or lean on a pile of references. It leads Artificial Analysis in image to video and sits in a five-way tie at the top of text to video.
Budget for 720p and $0.10 to $0.47 a second, and treat the 4K film variant as a trial product for now. Start on a plan from $25 a month and run Seedance 2.5 yourself.