Seedance 2.5 Pricing, Access and How ByteDance's Model Works
Seedance 2.5 explained: ByteDance's 30-second AI video model, its API pricing, where to use it, and how it compares with Kling, MiniMax and Veo.
Jordan Wells covers startups, applied AI, and the people building them.
ByteDance's Seed team officially launched Seedance 2.5 on July 31, 2026, and the headline number is length. The model generates up to 30 seconds of video with synchronized audio in a single pass, double the 15-second ceiling of the Seedance 2.0 family it replaces. Nearly three months later, the model has moved from ByteDance's own apps into a paid developer API and onto third-party creative platforms, which makes it a good moment to look at what it does, what it costs, and how it compares with the other flagship video generators on the market.
What Seedance 2.5 is
Seedance 2.5 is an audio-video joint generation model. It produces picture and sound together rather than adding a soundtrack afterward. In its official Seedance 2.5 launch post, ByteDance Seed groups the upgrade into three areas.
Long-form storytelling. A single generation can run to 30 seconds and can be extended over multiple rounds. ByteDance says the model can organize several connected shots inside one clip, so a story moves through setup, development and resolution instead of stretching one moment. Extension keeps characters, environments and pacing consistent, which the company says lets users build multi-minute videos without repeatedly splicing clips.
Multimodal references. Users can supply up to 30 images, 10 video clips and 10 audio clips as reference material in one request, 50 assets in total. The previous generation allowed 9 images, 3 videos and 3 audio clips.
Editing. Seedance 2.5 adds timestamp-level control for targeted edits to audio and video, plus green screen, camera perspective and reference-based editing aimed at film and advertising work.
The BytePlus ModelArk developer tutorial fills in the technical specification. The API model ID is dreamina-seedance-2-5-260628. Output duration is any value from 4 to 30 seconds, or -1 to let the model choose. Output resolution is 480p, 720p or 1080p, with 1080p delivered at 10-bit color depth. Supported aspect ratios are 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, and output can be MP4 or a new MOV format intended for editing workflows. The tutorial also lists native support for 11 languages, including English, Chinese, Spanish, Arabic, Japanese and Korean.
One detail is easy to miss. Seedance 2.0 offered a 4K output tier on the API, while Seedance 2.5 tops out at 1080p. In this release, the API's resolution ceiling moved down from 4K to 1080p even as clip length doubled.
How Seedance 2.5 compares with other flagship models
To place the 30-second figure in context, Stanford Tech Review compared the maximum single-generation clip length and maximum listed output resolution of current flagship video models, using only each vendor's own documentation.
Across four flagship video models whose official spec pages we checked on October 23, 2026, Seedance 2.5's 30-second ceiling is the longest single generation, twice that of Kling 3.0 and MiniMax H3 (15 seconds) and 3.75 times Google's Veo 3.1 (8 seconds), yet its 1080p maximum is the lowest output resolution of the four.
| Model | Developer | Max single-generation length | Max listed output resolution | Official source |
|---|---|---|---|---|
| Seedance 2.5 | ByteDance | 30 seconds | 1080p | BytePlus ModelArk tutorial |
| Kling 3.0 / 3.0 Omni | Kuaishou (Kling AI) | 15 seconds | 4K | Kling AI API documentation |
| MiniMax H3 | MiniMax | 15 seconds | 2K | MiniMax platform documentation |
Veo 3.1 (veo-3.1-generate-001) |
8 seconds | 4K | Google Cloud model documentation |
Source: vendor documentation (BytePlus ModelArk, Kling AI, MiniMax, Google Cloud), checked Oct 23, 2026. Excludes extension features.
Method: lengths are the largest duration value each vendor's API documentation accepts for one generation, before any extension feature. Kling's text-to-video API reference accepts integer durations from 3 to 15 seconds and resolutions up to 4K. MiniMax's video generation guide lists 4 to 15 seconds and 768P or 2K output for MiniMax H3. Google Cloud's Veo 3.1 page lists clips of 4, 6 or 8 seconds and output at 720p, 1080p or 4K for the GA model.
Several models were left out. OpenAI's Sora 2 is gone from the comparison because, according to OpenAI's Sora 2 model page, the Sora 2 models and Videos API were shut down on September 24, 2026. Google's newer Gemini Omni Flash and Runway's Gen-4.5 were excluded because their official pages we could access did not state a maximum single-generation length. Kling has also announced early access to a Kling 4.0 API; it is not in the table because its specifications were not part of the documentation we reviewed.
Two caveats apply. Extension features change the picture for longer projects: Seedance 2.5 supports multi-round extension, and Google's Omni Flash documentation describes extending clips by 10 seconds up to 40 seconds total. Resolution labels are also not always like-for-like, since some vendors note that higher tiers are upscaled.
Seedance 2.5 is the model to pick when a scene has to hold together for half a minute, and the wrong one when the deliverable is a sharp 4K frame.
How much Seedance 2.5 costs
Pricing depends on where you use it.
BytePlus ModelArk API. ByteDance's international cloud bills Seedance 2.5 by token, not by second. The ModelArk pricing page lists $10.70 per million tokens for 480p and 720p output and $11.70 per million tokens for 1080p, for requests without video input. Token use is estimated as input video duration plus output duration, multiplied by width, height and frame rate, divided by 1,024. BytePlus's own worked examples for a 5-second 16:9 clip without video input are $0.514 at 480p, $1.156 at 720p and $2.843 at 1080p, roughly $0.10, $0.23 and $0.57 per second. Applying the same published formula to a full 30-second 1080p clip at 24 frames per second gives an estimate of about $17. Requests that include video input carry minimum token charges, and BytePlus recommends an account balance or savings plan of at least $30 before enabling the model. Failed generations are not charged.
Runway API. Runway lists Seedance 2.5 in its developer model catalog as "Cinematic video up to 30 seconds at 480p, 720p, and 1080p," priced from $0.20 per second, with draft mode available.
Consumer and creator apps. Pricing in subscription apps is set by each platform in credits or membership tiers, so the per-clip cost varies.
Where you can use Seedance 2.5
At launch, ByteDance rolled the model out on its own Jimeng AI and Doubao Pro apps. The Seed blog gives the path as Jimeng Web, then Video Generation, then Seedance 2.5, with the same flow in Doubao Pro. API access, which ByteDance described as coming soon at launch, is now live on BytePlus ModelArk.
Outside ByteDance, the model is available through Runway's API and through creative platforms that license it. One example is OiiOii, which offers Seedance 2.5 AI video generation to members. Its product page lists clips of 4 to 30 seconds in one continuous shot, 1080p output, up to 50 reference inputs across images, audio and video, audio generated by default with an option to turn it off, and area editing of a selected region. It positions the model for anime, manga and live-action styles, and its paid plans start at a $16 monthly equivalent on quarterly billing.
How to use Seedance 2.5
Whichever platform you choose, the workflow follows the same steps.
1. Pick a generation mode
The API supports text-to-video, first-frame and first-and-last-frame video, multimodal reference generation, editing and extension. Reference mode is where Seedance 2.5 differs most from its predecessor, because it accepts audio-only references as well as image and video.
2. Write the prompt as a shot list
ByteDance's own examples read like a director's notes, with camera moves and time ranges such as "0 to 5s" and "6 to 10s." Because a 30-second clip can contain several shots, describing each beat in order gives the model a structure to follow.
3. Use draft mode to save money
On the API, draft mode generates a 480p preview at the 480p token rate. Once the preview looks right, its task ID can be reused to render the final 1080p version. The two steps are billed separately, but a cheap preview reduces the number of expensive full-resolution failures.
4. Extend instead of restarting
For anything longer than 30 seconds, extend the existing clip rather than generating a new one, so characters and settings stay consistent across segments.
The bottom line
Seedance 2.5 moves ByteDance to the front of the field on single-clip length and reference capacity, while conceding resolution to Kling, MiniMax and Google. For storyboard-driven work, advertising cuts and narrative shorts, that trade will often be worth it. For projects where final image sharpness matters most, the 1080p ceiling is the limit to plan around.