Seedance 2.5: ByteDance Takes AI Video to 30-Second Single Takes

ByteDance released Seedance 2.5, the newest version of its video generation model, on 31 July 2026. The headline change is duration: the model produces up to 30 seconds of video in a single pass, double what the previous version managed, with audio generated at the same time rather than added afterwards.

The release matters because clip length has been the practical ceiling on AI video. Most competing models produce five to ten seconds. That is enough for a social post and not enough for a scene.

What ByteDance confirmed

ByteDance’s Seed research team published the launch details in a post titled “One-take Creation, Flexible Referencing”. The confirmed capabilities are:

Thirty seconds in one generation. The model outputs up to 30 seconds per request and supports multiple rounds of extension for longer sequences. ByteDance describes the extension workflow as appending shots while holding character and environment consistency.

Fifty reference inputs. A single request accepts up to 30 images, 10 video clips and 10 audio clips. That is a large jump from the previous generation and is the feature most likely to matter to commercial users, because it is how you keep a specific product, actor or logo stable across a sequence.

Timestamp-level editing. Users can target edits to a specific point in the clip for both audio and video, rather than regenerating the whole thing.

Additional editing modes. ByteDance lists green screen, camera perspective and reference-based editing, along with support for clay render, motion and creative references. The clay render option lets a creator block out rough 3D geometry and have the model treat it as spatial guidance for camera work.

Joint audio and video. Sound is produced in the same pass as the picture rather than dubbed on afterwards, which is intended to fix the lip-sync and timing drift that separate audio pipelines introduce.

Where it is available

At launch, Seedance 2.5 rolled out on Jimeng AI and the Pro tier of Doubao, both ByteDance consumer apps aimed primarily at the Chinese market. ByteDance said API access was coming through BytePlus ModelArk.

That API has since arrived. The BytePlus product page for Seedance now lists the Dreamina Seedance 2.5 API as available, with output at 480P and 720P and durations from 4 to 30 seconds. BytePlus sells access through three plan tiers Light, Production and Premium each described in terms of approximate seconds of 480P video rather than a flat per-second rate.

Consumer access outside China runs mainly through Dreamina, ByteDance’s CapCut-linked creative platform, and through independent web front-ends that call the model. Browser-based services such as Seedance 2.5 package text-to-video, image-to-video and reference-guided editing in a single interface with subscription plans. These are third-party products rather than ByteDance-operated ones, and buyers should confirm output resolution, watermark policy and commercial licensing before committing.

What has not been confirmed

Several widely repeated specifications are not backed by ByteDance’s technical launch post, and buyers should treat them carefully.

Resolution is the clearest example. The Seed launch post does not state a maximum output resolution or a frame rate for Seedance 2.5. The BytePlus API listing specifies 480P and 720P. Separately, ByteDance’s Dreamina product page markets 4K output and up to 60 fps as premium features, and describes a beta long-video mode extending clips to 180 seconds. Those are product-page marketing claims tied to a specific consumer surface, not model specifications published by the research team, and the API tier available to developers currently tops out lower.

Pricing is the second gap. ByteDance has not published a public per-second rate for Seedance 2.5. Older pricing guidance covered the Seedance 2.0 family, where the mini variant was billed at 40 percent of the standard rate and the fast variant at 75 percent. Applying those numbers to 2.5 would be guesswork.

Independent benchmarking is the third. There was no third-party leaderboard placement for Seedance 2.5 at launch. The German outlet The Decoder reported that Artificial Analysis had ranked the previous version, Seedance 2.0, at the top of its image-to-video leaderboard among models with audio a useful signal about the family, but not a measurement of 2.5.

How it compares

The most direct comparison is Google’s Gemini Omni Flash, announced on 30 June 2026, which generates video from text, image and video inputs and supports conversational editing. Google’s model currently produces clips of up to 10 seconds, with longer durations described as coming, and is priced at $0.10 per second of output.

On duration, Seedance 2.5 is well ahead. On transparency, Google is ahead: a published per-second price and a documented API make cost modelling straightforward, which is what enterprise buyers need before they commit to a pipeline.

The practical read

For creative teams, the useful question is not which model wins a demo reel. It is which one survives a revision cycle.

Seedance 2.5’s argument is that it does. Thirty seconds means a full scene rather than a fragment. Fifty reference inputs means the client’s product can look like the client’s product. Timestamp-level editing means a note like “change the line at 12 seconds” does not force a full regeneration. Those are workflow features, and they are aimed at agencies and marketing teams rather than at people making novelty clips.

The counterweight is that the strongest surfaces are still China-first consumer apps, the developer API launched at lower resolutions than the marketing suggests, and there is no published rate card to plan a budget against. Teams outside China evaluating Seedance 2.5 should run a paid pilot on the BytePlus tiers and measure real cost per finished second before designing a production process around it.