The AI arena is free today

Open Superagent
Back to blog
Model Release·Video

Seedance 2.5: ByteDance's 30-Second One-Take Video Model

Seedance 2.5 extends single-pass audio-video to 30s with denser references and timestamp edits. ModelArk: $10.70/$6.40 per 1M, 480p/720p only, not available in the United States.

Sebastian Crossa
Sebastian Crossa
Co-Founder @ LLM Stats
·4 min read
Seedance 2.5: ByteDance's 30-Second One-Take Video Model

Seedance 2.0 already did joint audio-video. 2.5 is ByteDance pushing from "generate a clip" toward "finish a piece": longer takes, multi-round extension, denser references, and edit passes that keep continuity.

At a Glance

  • Product release: July 31, 2026
  • API: BytePlus ModelArk ~August 7, 2026 (not available in the United States per BytePlus; may need Dreamina Seedance 2.5 resource pack to enable)
  • Model ID (ModelArk): dreamina-seedance-2-5-260628 (also surfaced as doubao-seedance-2.5 on Volcano Engine Ark in China)
  • Organization: ByteDance Seed
  • API duration: [4,30] seconds or -1; multi-round extensions for longer workflows
  • References (single pass): Up to 30 images, 10 videos, 10 audio clips; total ref video ≤30s and total ref audio ≤30s; pure audio-only refs are new vs 2.0
  • Resolutions (ModelArk API): 480p and 720p only (default 720p). Dreamina consumer claims of 4K / ~3min long-video are not the ModelArk API
  • Consumer access: Jimeng AI, Doubao Pro (Video Generation → Seedance 2.5)
  • API pricing: $10.70 / 1M tokens (no video input); $6.40 / 1M (with video input)

What's Actually New vs Seedance 2.0

1. Thirty-second one-takes (and extend)

Official blog: single-pass generation moves from 15s to 30s. ModelArk duration is [4,30] or -1, with multi-round extension that tries to keep characters, environment, and pacing consistent.

2. Heavier multimodal referencing

One pass can ingest up to 30 images, 10 videos, and 10 audio clips, with total reference video and audio each capped at ≤30s. Pure audio-only references are new vs 2.0. Clay-render, motion, and creative reference modes cover blocking, camera paths, and style.

3. Timestamp-level editing

Prompts can target narrative, camera, and motion in specific time ranges. Post-generation edits can change characters, actions, or plot while aiming to preserve continuity.

4. Price, resolution, and region trade

Published ModelArk rates put 2.5 well above 2.0 per token (~53% without video input). API resolutions are 480p/720p only. BytePlus states the API is not available in the United States. Example five-second 16:9 costs (no video input): $0.514 at 480p, $1.156 at 720p. Verify usage.completion_tokens on real responses.

Specs & Pricing

DetailValueNotes
API duration[4,30] or -1Multi-round extension for longer
AudioNative joint audio-videoPure audio-only refs new vs 2.0
Reference caps30 images / 10 videos / 10 audioTotal ref video ≤30s; total ref audio ≤30s
ModelArk IDdreamina-seedance-2-5-260628Confirm in console / region
Resolutions480p, 720p (default 720p)No 1080p/4K on ModelArk API
RegionNot available in the United StatesPer BytePlus; Dreamina pack may be required
Token rate (no video in)$10.70 / 1M480p & 720p
Token rate (video in)$6.40 / 1MInput duration still costs tokens
Example 5s 480p$0.514No video input
Example 5s 720p$1.156No video input

Docs: https://docs.byteplus.com/en/docs/ModelArk/2607688 (tutorial), https://docs.byteplus.com/en/docs/ModelArk/1544106 (pricing). Re-check live docs before budgeting.

Benchmarks

No independent LLM Stats video arena numbers required for this launch brief. Do not carry forward Seedance 2.0 third-party leadership notes onto 2.5 without a fresh 2.5 measurement. Treat ByteDance demo quality claims as self-reported.

ByteDance's own caveats: room to improve on physical plausibility of complex motion and stability with many interacting subjects.

When to Use It

Good fit: Story-shaped shorts, concert / narrative one-takes, reference-heavy multi-character scenes, education and industrial synthetic video, edit-and-extend workflows outside US ModelArk restrictions.

Watch the bill and region: Long reference video/audio can inflate tokens (≤30s total each). US users may need consumer Dreamina paths instead of ModelArk. Compare cost-per-accepted-clip against Seedance 2.0.

Not magic yet: Complex multi-subject physics and long interaction scenes are where the vendor says quality still slips. Do not assume Dreamina 4K / ~3min marketing applies to the API.

Outlook

Seedance 2.5 is a workflow release: longer one-takes, denser references, timestamp edits, with API caveats on region, ref duration caps, and 480p/720p-only ModelArk resolution. The useful comparison is cost per accepted story beat on your prompts and region, not sticker rate or consumer 4K claims alone.

Sources: https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5 , https://seed.bytedance.com/en/seedance2_5 , https://docs.byteplus.com/en/docs/ModelArk/2607688 , https://docs.byteplus.com/en/docs/ModelArk/1544106

Questions

Frequently Asked Questions

  • Product platforms July 31, 2026; ModelArk API around August 7, 2026.

  • ModelArk duration [4,30] or -1, with multi-round extensions for longer pieces.
  • Published ModelArk rates: $10.70 / 1M tokens without video input, $6.40 / 1M with video input (480p / 720p). Example 5s clips: $0.514 (480p) and $1.156 (720p) without video input.

  • dreamina-seedance-2-5-260628 on BytePlus ModelArk (verify in your region; not available in the United States).
  • Jimeng AI and Doubao Pro video generation (Dreamina marketing may advertise higher resolution / longer video than ModelArk).

Continue Reading