What Is Seedance 2.5? Features, Availability & How to Access

Seedance is ByteDance's family of AI video generation models — the lineage behind Seedance 1.0, the audio-capable 1.5 Pro, and the widely used Seedance 2.0. The newest step is Seedance 2.5, and it is no longer a preview: you can generate with it today.

Is Seedance 2.5 out yet?

Yes. Seedance 2.5 was officially unveiled on June 23, 2026 at the Volcano Engine FORCE conference and launched into a closed (enterprise) beta. Through July it stayed there — the widely reported July 16 API opening came and went with no public API, model listing, or pricing. That has since changed. Seedance 2.5 is now live across the board:

• Sunra — generally available as of August 2026, and the route this site links to.
• Dreamina / CapCut — Dreamina's model list carries Seedance 2.5 as its newest model.
• Volcano Engine Ark — model doubao-seedance-2-5-260628, documented and callable.
• BytePlus ModelArk — model dreamina-seedance-2-5-260628, the international API surface.
• fal and Replicate — both publish bytedance/seedance-2.5 endpoints.

One honesty note on dates: ByteDance never posted a single dated "Seedance 2.5 is generally available" announcement. The closest dated first-party marker is CapCut's July 31, 2026 newsroom post announcing the Dreamina Seedance 2.5 rollout to subscriber accounts across Europe, Asia, the Middle East and South America. So treat any precise "Seedance 2.5 launch day" figure you see elsewhere as unverified. For the fuller history, see the Seedance 2.5 release date.

What Seedance 2.5 adds

The upgrades over 2.0 change what's actually practical to make:

• Clips of 4 to 30 seconds in a single generation — double 2.0's ~15-second ceiling, with no stitching. A longer native clip means a full scene that builds, holds, and resolves, instead of a short loop with visible seams where separate generations were joined.
• Up to 50 reference inputs in one prompt — 30 images, 10 videos and 10 audio clips, a big jump from 2.0's 9 + 3 + 3. Reference video is capped at 30 seconds in total across those files, and audio-only references work on their own. More references means tighter control over characters, style, and continuity across a shot.
• Audio generated with the picture, in the same pass — dialogue, effects and ambience arriving already in sync rather than laid on afterwards.
• 480p, 720p or 1080p output (720p by default), in 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16, plus an adaptive mode that follows your reference. You can ask for up to four clips per run. Worth knowing: not every route exposes every tier — fal, for instance, publishes only 480p and 720p — so check the platform before you plan on 1080p.

ByteDance also announced localized, controllable editing — changing part of a shot without regenerating the whole clip. That one we'd flag as an announcement claim rather than something you can confirm from the generation routes today, so don't plan a workflow around it yet.

And one correction worth making, because many pages get it wrong: native 4K was a Seedance 2.0 update, not a 2.5 feature. Seedance 2.5 tops out at 1080p, so "30 seconds and 4K" is not a combination it offers.

How to access it

• Sunra — the zero-setup route, and the one this site links to. Seedance 2.5 is live there as a subscription model; pick it, paste a prompt, generate.
• Dreamina / CapCut — ByteDance's consumer creation platforms, where 2.5 sits at the top of the model list. Jimeng (the China app) typically needs a Chinese phone number; Dreamina international does not.
• Volcano Engine (Ark) API — ByteDance's direct, programmatic route in China, aimed at businesses and pipelines.
• BytePlus ModelArk — the international API surface.
• fal and Replicate — authorized partners with per-second REST endpoints.

For a route-by-route walkthrough, see how to access Seedance 2.5.

What's still unconfirmed

ByteDance has not published a first-party public rate card for 2.5, so pricing depends entirely on the platform you use — on Sunra it bills in credits that scale with clip duration and resolution. Early pricing chatter extrapolated from Seedance 2.0 (roughly 1 RMB per second of video) is a rough order of magnitude at best, not official. Third-party reports also describe a Dreamina long-video mode stretching 30-second scenes toward 90- and 180-second outputs; nothing we can verify generates past 30 seconds in one pass. Some coverage has raised content-ownership and copyright questions around long AI clips; expect the usual platform content policies to apply.

How to get a good first result

Prompt skills carry straight over between Seedance versions, so the fastest start is a prompt that already works rather than a blank box:

1. Copy one that's proven — browse Seedance 2.5 prompts and take a few.
2. Learn the structure Seedance responds to — see how to use Seedance.
3. Run it on Sunra, swap the subject for yours, and keep the outputs you like.

Because 2.5 will happily give you 30 seconds, it rewards prompts that describe a shot developing — a beginning, a movement, a resolution — rather than a single frozen image. For how it compares feature-by-feature, see Seedance 2.5 vs 2.0.