Seedance 2.5 API: 30-Second Video, Audio & How to Access

Seedance 2.5 is ByteDance's next video model, unveiled at the Volcano Engine FORCE conference on June 23, 2026. For anyone building with the API, three changes matter: a video call can now return a native 30-second clip in one shot (2.0 capped out around 15s), a single request can carry up to 50 joined reference assets, and the model can generate synced audio (voice, sound effects, and background music) in the same pass.

If you already ship on Seedance 2.0, treat 2.5 as a drop-in for the jobs where 2.0 forced you to stitch clips together or bolt on audio afterward. The API surface is close enough that a migration is mostly raising a few limits and adding two or three new fields.

The capabilities and limits below come from ByteDance's Seedance 2.5 integration documentation and the June 23 FORCE announcement. As of July 2026 the public API is rolling out in stages, so the one thing that varies by where you call it is the final route ID and price. Confirm those on your provider; everything else is the model's actual behavior.

What Seedance 2.5 adds over 2.0

The jump is about length, references, and sound — not a cosmetic quality bump.

• 30-second native generation. duration now accepts [4, 30] seconds (roughly 97–721 frames) instead of topping out near 15s. A 30-second product demo, with setup, action, and payoff, is one task instead of a stitch-and-pray pipeline.

• Up to 50 joined references. A single request accepts 30 images + 10 videos + 10 audio clips (2.0 allowed 9 / 3 / 3). Reference videos and audio can each run up to 30s per clip and 30s total per modality. This is the number most third-party pages get wrong; some still quote 20 images / 6 clips.

• Native audio. With generate_audio on, the output ships with dialogue, sound effects, and BGM matched to the visuals. It builds on the audio-visual sync work in Seedance 1.5 Pro, now paired with 30-second range.

• Better editing and control. Multi-pass video extension no longer smears — extending a clip three times keeps it sharp. Negative instructions like "no subtitles, no BGM" are followed more reliably, timestamped prompts land more precisely, multilingual scenes hold up, and duplicate-person ("multiple twins") artifacts drop.

• A professional color path. 2.5 adds mov output with yuv444 color, which survives repeated re-encoding during extension and grading far better than the default mp4.

For a full side-by-side and a "should I switch?" verdict, see Seedance 2.5 vs Seedance 2.0.

Is Seedance 2.5 available yet? Release date and access

Seedance 2.5 was announced on June 23, 2026 at Volcano Engine FORCE in Beijing. API access is rolling out through mid-2026, starting with staged and partner access rather than a single open switch. That's why third-party pages vary so much on "release date": the announcement is fixed, but the open API window is staged.

There are three realistic ways to reach it:

1. First-party — ByteDance's Volcano Engine / BytePlus ModelArk platform, opening in stages.

2. Third-party model routers — aggregators that resell the route once it's live.

3. AIReiter — we add new ByteDance video models as soon as they ship (as we did across the Seedance 2.0 line), so a Seedance 2.5 route will land here the moment it's generally available, behind the same OpenAI-compatible endpoint you already use.

Whichever route you pick, confirm the final model ID and current price on that provider before you wire it into production.

How to call the Seedance 2.5 API

Seedance 2.5 uses the same async, content-array shape as the rest of the Seedance line: you submit a job, get a task ID back, then poll or receive a callback when the video is ready.

A minimal image-to-video request with native audio and a 30-second target looks like this (POST it to your provider's video-generation endpoint with your key in the Authorization: Bearer <API_KEY> header):

{
  "model": "doubao-seedance-2-5-260628",
  "content": [
    { "type": "text", "text": "A barista pulls an espresso shot, steam rising, morning light. \"Best cup all week,\" she says." },
    { "type": "image_url", "role": "first_frame", "image_url": { "url": "https://your-cdn.com/cafe.jpg" } }
  ],
  "duration": 30,
  "resolution": "720p",
  "ratio": "16:9",
  "generate_audio": true,
  "output_format": "mp4",
  "callback_url": "https://your-app.com/webhooks/seedance"
}

The fields most requests touch:

Because a 30-second job runs longer than an image job, lean on the callback instead of tight polling, and log every state — queued and running clips will sit in the queue longer than you're used to from 2.0.

Getting the most from 30 seconds, 50 references, and native audio

• Plan for one shot, not a stitch. With 30 seconds you can hold a full beat — a camera move and reveal, or a three-part action, in a single generation, which kills the continuity drift you'd get from joining 15-second clips.

• Spend the reference budget deliberately. 30 images / 10 videos / 10 audio is a lot of room. Use images for characters, products, and style; use reference videos for motion and pose; use reference audio for a voice or rhythm you want matched. Validate both the per-asset limit and the per-modality total — a request can pass the single-file check and still fail because total reference video duration exceeds 30s.

• Put spoken lines in double quotes. For cleaner voice sync, wrap dialogue in quotation marks inside the prompt (e.g. she says, "we're closed"). The model treats quoted text as lines to voice rather than scene description.

mp4 vs mov: which output format

Most delivery workflows should stay on mp4 — it's the smaller, universally playable default (H.264 / AAC / yuv420p), ready to drop straight onto web, mobile, and social.

Reach for mov when the generated clip is an intermediate, not the final deliverable. mov carries yuv444 color and PCM audio, which hold up through extension, compositing, and color grading where mp4's chroma subsampling would drift over repeated encodes. The trade-off is a much larger file. Rule of thumb: deliver in mp4, edit in mov. (mov is a 2.5-only option; older models reject it.)

Draft mode: preview cheap, then render the final

If you're iterating on a prompt, don't burn full-resolution 30-second renders on drafts. Seedance 2.5's draft mode generates a fast 480p preview and returns a draft_task ID. Once a preview looks right, you pass that ID back to reuse the exact inputs (prompt, references, ratio, duration) and render the finished clip. It's the cheapest way to nail a shot before you pay for the full-quality pass, and the single biggest cost lever most people miss.

Seedance 2.5 pricing and cost

There's no single public flat rate yet; price depends on your provider and settings. What drives cost is predictable:

• Duration. A 30-second clip costs meaningfully more than a 5- or 15-second one; you're paying per frame.

• Resolution. 720p costs more than 480p.

• References and audio. Heavier multimodal inputs and generate_audio add processing.

There is no free tier. The closest thing to free experimentation is draft mode, which lets you preview at 480p before committing to a full render. If you route through AIReiter, Seedance 2.5 sits alongside the rest of your models on one balance, so a team already running Seedance 2.0 cost optimization can extend the same tactics — draft-first, right-size resolution, reuse last frames — without a second billing account. Weighing it against Sora 2, Veo 3.1, or Kling? See the 2026 AI video model comparison.

FAQ

When was Seedance 2.5 released?

It was announced on June 23, 2026 at Volcano Engine FORCE. API access is rolling out in stages through mid-2026 rather than launching all at once.

Is there a free version of Seedance 2.5?

No free tier. Draft mode (a fast 480p preview) is the cheapest way to test a prompt before paying for a full render.

How long can Seedance 2.5 videos be?

Up to 30 seconds in a single native generation, double the ~15-second ceiling of Seedance 2.0. The duration range is [4, 30].

Does Seedance 2.5 generate audio?

Yes. With generate_audio enabled it produces synced voice, sound effects, and background music. Put spoken lines in double quotes for better sync.

How many reference assets can Seedance 2.5 take?

Up to 50 total: 30 images, 10 videos, and 10 audio clips per request, with reference video and audio capped at 30s per clip and 30s total each.

Does Seedance 2.5 support 4K?

Not at launch. The launch build outputs 480p and 720p; higher resolutions may come later.

Is Seedance 2.5 better than 2.0?

For long, reference-heavy, or edit-heavy work, yes. For short clips that already work on 2.0, the upgrade is optional. Full breakdown: Seedance 2.5 vs Seedance 2.0.

How do I access the Seedance 2.5 API?

Through ByteDance's Volcano Engine / BytePlus platform as it opens in stages, through model routers, or via AIReiter, which adds new ByteDance video models on release behind an OpenAI-compatible endpoint.