Seedance 2.5 is ByteDance's next-generation AI video model, and it's now live: single-pass clips up to 30 seconds, up to 50 multimodal references in one generation, resolution up to 720p, and audio that's generated together with the picture instead of added afterward. This guide covers what it is, what changed compared to Seedance 2.0, how pricing and access work, and how to get started. 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
Seedance 2.5 is live. ByteDance announced it on June 23, 2026, at the Volcano Engine FORCE conference in Beijing, opened broader API access through BytePlus in mid-July, and the model launched July 31, 2026. Apiframe added Seedance 2.5 to its unified API on August 7, 2026, so it's reachable today through the same endpoint, key, and billing as every other model on the platform.
One correction worth flagging: ByteDance's original announcement promised native 4K output, but what actually shipped tops out at 720p (480p and 720p are the two available resolutions). We've corrected the resolution claims throughout this guide to match what's actually available via the API rather than the pre-launch announcement. Everything else ByteDance announced, 30-second single-pass clips, up to 50 multimodal references, native synced audio, region-level editing, and director-style camera control, held up as advertised.
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.
Up to 720p resolution with synchronized audio
Seedance 2.5 renders at up to 720p (480p and 720p are the available options). That's lower than what ByteDance's original announcement promised (native 4K, 10-bit color), a discrepancy worth flagging since it's the kind of thing that trips people up if they build around the announcement instead of the shipped spec. 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 | 4–30 seconds, in one continuous pass |
| Reference inputs | Up to around 12 | Up to 50 (images, video, and audio combined) |
| Resolution | Up to 1080p | 480p or 720p (ByteDance's announcement promised native 4K; that didn't ship) |
| 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 |
This table originally reflected ByteDance's pre-launch announcement. It's now updated to match the specs Seedance 2.5 actually shipped with, resolution is the one meaningful correction. We'll keep this current as more real-world testing comes in.
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's confirmed and live now that the model has shipped.
ByteDance still hasn't published a specific per-second rate for Seedance 2.5 the way it has for Seedance 2.0 (around $0.06/second on standard tiers, roughly $0.022/second on faster, lower-cost tiers through third-party providers). Since Seedance 2.5's actual resolution ceiling (720p) is lower than Seedance 2.0's (1080p), it's no longer safe to assume the newer model costs more per second, pricing could land anywhere depending on how ByteDance weighs the longer duration and higher reference count against the lower resolution. Check the Seedance 2.5 API page for current, confirmed pricing.
How to access Seedance 2.5
There are two practical routes, and both are live now.
The official route runs through BytePlus ModelArk or Volcano Engine directly. 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 live now, 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 start generating today, 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, confirmed now that the model has shipped
- 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 and launched July 31, 2026. Its headline features are single-pass 30-second clips, up to 50 multimodal references, up to 720p resolution, region-level local editing, director-grade camera control, and native synchronized audio.
Is Seedance 2.5 out yet?
Yes. Seedance 2.5 launched July 31, 2026, and Apiframe added it to its unified API on August 7, 2026.
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), and editing (full regeneration to region-level local editing). Resolution actually moved the other way: Seedance 2.0 tops out at 1080p, while Seedance 2.5 currently caps at 720p, ByteDance's original 4K announcement didn't make it into the shipped model. 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: 15 credits ($0.15) per second at 480p, 34 credits ($0.34) per second at 720p. Check the Seedance 2.5 API page for current pricing.
How do I get API access?
Through BytePlus ModelArk or Volcano Engine directly, or through a unified API like Apiframe, which gives you a single key and endpoint without a separate ByteDance account. Both routes are live now.
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.