Seedance 2.5 is ByteDance's next-generation AI video model, and it's built around four headline numbers: single-pass clips up to 30 seconds, up to 50 multimodal references in one generation, native 4K output, and audio that's generated together with the picture instead of added afterward. This guide covers what it is, what's actually new compared to Seedance 2.0, how pricing and access work, and how to get started. We keep it updated as the model finishes rolling out. If you land here from a comparison post, a tutorial, or a search for "Seedance 2.5," this is the page those all point back to.
Current status
As of this update, Seedance 2.5 is not yet generally available. ByteDance announced it on June 23, 2026, at the Volcano Engine FORCE conference in Beijing, and it's currently limited to a closed enterprise beta. There's no public sign-up and no officially confirmed release date. The widely expected window is early July 2026, based on how ByteDance has talked about the rollout, but treat that as an estimate rather than a promise.
The rollout is happening in stages rather than all at once. Seedance 2.5 is expected to appear first through Doubao and Volcano Engine in China, with international access following through BytePlus ModelArk and the Dreamina app. Pricing for Seedance 2.5 specifically hasn't been published yet either. We'll update this section and the rest of the guide the moment that changes.
What is Seedance 2.5?
Seedance 2.5 is ByteDance's next-generation multimodal AI video model, and it's a bigger jump than the version number implies. The company skipped straight from 2.0 to 2.5, which usually signals a genuinely different set of capabilities rather than an incremental point release, and the feature list backs that up.
The model is served through Doubao and Volcano Engine in China, and through BytePlus ModelArk and the Dreamina app internationally. Once it's reachable through an API, it supports text-to-video, image-to-video, and reference-to-video generation behind a single endpoint, with an async job model: you submit a request, then poll or receive a webhook when the result is ready.
What's new: full capability breakdown
Single-pass 30-second clips
Most longer AI-generated videos today are actually several shorter clips stitched together, and stitching introduces small inconsistencies at every seam: a character's face drifts slightly, lighting jumps, and motion stutters. Seedance 2.5 generates a single continuous clip up to 30 seconds in one pass, with scene changes and tempo shifts built into that same generation, so there's no seam for those problems to appear at in the first place.
Up to 50 multimodal references
A single generation can take up to 50 multimodal references, combining images, short video clips, and audio. That's a substantial jump from around 12 on the previous generation. Each reference can be tagged with a role, so the model knows exactly which input should pin down a character's identity, a product's look, a brand's color palette, a camera style, or a voice.
Native 4K with synchronized audio
Seedance 2.5 renders natively at up to 4K (3840x2160) with 10-bit color, which means more headroom for color grading and post-production than an upscaled clip would give you. Audio is co-generated in the same pass as the video rather than dubbed on afterward, so music, sound effects, dialogue, and lip-sync land synchronized to the picture by default.
Director-grade camera and multi-shot control
You can describe multi-shot sequences and camera direction (pans, dollies, orbits, focus changes) directly in the prompt, and the model keeps the subject consistent across those cuts. ByteDance reports roughly 20% better prompt adherence than Seedance 2.0, which shows up most clearly in how reliably the camera does what you actually asked for.
Region-level local editing
You can edit a specific region or element of a clip while the rest of the frame stays visually stable, or use an existing clip as a reference and change only what you specify, keeping its character, environment, camera movement, composition, action rhythm, and duration intact. That means fixing one detail no longer requires a full re-roll of the whole generation.
3D white-model previsualization
Seedance 2.5 also supports blocking out a scene in a rough 3D white-model pass before committing to a full render, letting you lock in layout, framing, and motion earlier in the process.
Seedance 2.0 vs Seedance 2.5
ByteDance has been fairly specific about what changed. Here's the comparison based on the company's own announced specs:
| Capability | Seedance 2.0 | Seedance 2.5 |
|---|---|---|
| Max clip length | Around 15 seconds | 30 seconds, in one continuous pass |
| Reference inputs | Up to around 12 | Up to 50 (images, video, and audio combined) |
| Resolution | Up to 1080p | Native 4K (3840x2160), 10-bit color |
| Audio | Native synced audio, over shorter clips | Native synced audio, sustained across the full 30 seconds |
| Editing | Full regeneration to change any part of a clip | Region-level local editing, change one part and keep the rest |
| Camera control | Prompt-based, more limited | Director-style multi-shot control (pans, dollies, orbits) |
| Prompt adherence | Baseline | About 20% better, per ByteDance |
These are ByteDance's own promised specs from the announcement, not independently benchmarked numbers. We'll revisit this table with real, tested figures once Seedance 2.5 is out and people (including us) have had a chance to put it through its paces.
Pricing
Seedance 2.5 uses a per-second billing model: cost scales with duration, resolution, and the number of references you use, and you only pay for successful generations. That mechanism is confirmed. What isn't confirmed yet is the actual per-second rate for Seedance 2.5 specifically, since ByteDance hasn't published official pricing.
The only current reference point is Seedance 2.0, which runs around $0.06 per second on standard tiers and roughly $0.022 per second on faster, lower-cost tiers through third-party providers. Since native 4K, 10-bit output and 30-second single-pass generation require meaningfully more compute than Seedance 2.0's shorter, lower-resolution clips, it's reasonable to expect a higher rate or resolution-based pricing tiers, but we're not going to publish a specific number until ByteDance does. Check the Seedance 2.5 API page for current, confirmed pricing.
How to access Seedance 2.5
There are two practical routes.
The official route runs through BytePlus ModelArk or Volcano Engine directly, which currently means enterprise beta account setup and approval. It's the right path if you're already inside ByteDance's ecosystem or need a direct enterprise relationship.
The faster route is a unified API: one key, no separate ByteDance or BytePlus account required, with async jobs and webhooks built in from the start. Apiframe's Seedance 2.5 API is built to give you that single endpoint the moment the model is reachable, and since it shares billing and authentication across every model on the platform, you're not managing a separate account just for this one model. Get an API key to be ready the moment access opens up, or read the Seedance 2.5 API docs for the full request and parameter reference.
What Seedance 2.5 is good for
The combination of a 30-second single-pass narrative, a large reference budget, and native audio makes Seedance 2.5 particularly well suited to a specific set of use cases.
Long-form narrative and brand or character consistency benefit the most directly: holding a character's identity steady across movement, angle changes, and scene transitions matters for films, brand spots, and serialized content, and that's exactly what the reference system and single-pass generation are built for. Synced-audio content (dialogue, voiceover, music, and sound effects generated in the same pass as the picture) removes a whole separate production step. Product video at scale is another strong fit: a reference kit built from product photos, brand palette, and a voice sample can drive multiple on-brand variants without a film crew, and region-level editing lets you swap a label or color per market without regenerating the whole clip. Multilingual localization works the same way, since native audio and on-screen text can render across many languages in a single generation.
For a closer look at building product-video pipelines or comparing Seedance 2.5 against other video models on capability and cost, see Apiframe's AI video generator overview and use cases pages, or compare it directly against Veo 3.1 and other models on the full model list.
More on Seedance 2.5
This page is meant to be the starting point for everything Seedance 2.5 on Apiframe. A few related resources, some live now and some coming as the model rolls out further:
- Seedance 2.0 vs 2.5, a deeper migration-focused comparison for teams already building on 2.0
- Seedance 2.5 API pricing, with confirmed numbers once ByteDance publishes them
- A Seedance 2.5 quickstart tutorial, with working code in Python and JavaScript
- Seedance 2.5 for product video and brand content at scale
Until those go live individually, everything above is covered in the relevant sections of this guide.
FAQ
What is Seedance 2.5?
Seedance 2.5 is ByteDance's next-generation multimodal AI video model, announced June 23, 2026, with general availability expected in early July 2026. Its headline features are single-pass 30-second clips, up to 50 multimodal references, native 4K output, region-level local editing, director-grade camera control, and native synchronized audio.
Is Seedance 2.5 out yet?
Not publicly. As of this update, it's still in closed enterprise beta, with no confirmed public release date.
How is it different from Seedance 2.0?
The core differences are clip length (15s to 30s, in a single continuous pass), reference budget (about 12 to up to 50), resolution (1080p to native 4K with 10-bit color), and editing (full regeneration to region-level local editing). See the comparison table above for the full breakdown.
How much does the Seedance 2.5 API cost?
It's billed per second of generated video, but ByteDance hasn't published official Seedance 2.5 rates yet. Current cost estimates are based on Seedance 2.0 pricing, not confirmed 2.5 numbers.
How do I get API access?
Through BytePlus ModelArk or Volcano Engine's official enterprise beta, or through a unified API like Apiframe, which gives you a single key and endpoint without a separate ByteDance account.
Does it generate audio?
Yes. Audio is co-generated in the same pass as the video, producing native synchronized audio (music, sound effects, dialogue, and lip-sync) rather than a silent clip scored afterward.