AI DevelopmentNew Release6 min readPublished June 23, 2026

ByteDance FORCE 2026 · 30s native clip · enterprise beta now

Seedance 2.5: ByteDance's 30-Second AI Video Model

ByteDance announced Seedance 2.5 on June 23, 2026 at its Volcano Engine FORCE conference. It is not generally available yet — global enterprise beta now, with a public launch targeted for early July. The headline claims: a native 30-second clip, up to 50 reference inputs, and editing that re-draws part of a frame without touching the rest.

DA
Digital Applied Team
Senior strategists · Published June 23, 2026
PublishedJune 23, 2026
Read time6 min
SourcesByteDance, Artificial Analysis, +3
Announced
Jun 23
FORCE 2026
Native clip
30s
single pass
longest single-shot
Reference inputs
50max
~4x Seedance 2.0
Seedance 2.0 rank
#1
Artificial Analysis
ahead of Veo & Kling

ByteDance used its Volcano Engine FORCE conference on June 23, 2026 to announce Seedance 2.5, the next version of the text- and image-to-video model behind apps like Dreamina, Doubao, and CapCut. It is not generally available yet: Volcano Engine president Tan Dai framed it as a global enterprise beta, with a public launch targeted for early July 2026.

The headline claims are ambitious — a single native 30-second clip, up to 50 reference inputs in one generation, and editing that re-draws part of a frame without touching the rest. The reason to take them seriously is the version that already shipped: on independent blind-preference testing, ByteDance's current Seedance 2.0 sits at the top of the AI video field, ahead of Google and the major Chinese labs, at a noticeably lower price.

This guide separates what ByteDance has confirmed from what is so far only reported, places 2.5 on an accurate version timeline, and translates the announcement into what it actually changes for marketing and content teams. Where a figure is vendor-stated or carries a caveat, we say so.

Key takeaways
  1. 01
    Seedance 2.5 was announced June 23, 2026 — not yet shipped.ByteDance unveiled it at the Volcano Engine FORCE conference (president Tan Dai). It is in global enterprise beta now, with a public launch targeted for early July 2026.
  2. 02
    The marquee claim is a native 30-second single-pass clip.ByteDance says it generates a full 30 seconds in one native generation — pitched as the longest single-shot duration of any model — versus the roughly 5 to 15 seconds typical of rivals today.
  3. 03
    Up to 50 multimodal reference inputs, a ~4x jump.Images, audio, and video combined in one generation (Seedance 2.0 took 12) — the controllability story aimed squarely at holding characters, products, and style consistent.
  4. 04
    Local re-draw editing changes one element, keeps the rest.Swap a product, background, or subject without altering motion, camera, or lighting — demonstrated on the e-commerce problem of swapping a product into an existing clip.
  5. 05
    The credibility anchor is Seedance 2.0's current #1 ranking.On the independent Artificial Analysis Video Arena, the shipping Seedance 2.0 leads text-to-video and image-to-video, ahead of Google Veo 3.1 and Kling 3.0 — and at a fraction of their cost. No 2.5 benchmark exists yet.

01The AnnouncementByteDance ships 2.5 as a beta, not a same-day release.

The announcement came on stage at the 2026 Volcano Engine FORCE conference, ByteDance's annual cloud and AI event, attributed to Volcano Engine president Tan Dai. Notably, Seedance 2.5 was presented as a global enterprise beta with a public launch targeted for early July 2026 — a target to verify against ByteDance's own confirmation, not a fixed ship date. At the same event ByteDance also gave the current Seedance 2.0 a native 4K upgrade.

Status
Global enterprise beta
announced Jun 23, 2026 · GA targeted early July

Headline features: a native 30-second clip, up to 50 reference inputs, and local re-draw editing. All are vendor claims until independent testing at general availability.

verify GA date
Where it lives
One model, several apps
Model Ark API · Dreamina · Doubao · CapCut

ByteDance ships this family under several brands depending on the surface — Seedance, Dreamina/Jimeng, Doubao, and CapCut. Leaderboards often list it as 'Dreamina Seedance.' Same model, not separate products.

mind the naming
A note on the names
ByteDance's video family appears under several brands: Seedance (the Volcano Engine / Seed research name), Dreamina and Jimeng (the creative app), Doubao (the assistant), and CapCut. They are the same underlying model, not separate products — so coverage that names "Dreamina Seedance" or "Doubao video" is describing this same launch.

02Version TimelineFour named releases in roughly a year.

Placing 2.5 on an accurate timeline matters because the marketing around any single launch tends to blur which capabilities shipped when. For the full picture of the current model, see our Seedance 2.0 guide. One date worth pinning down: Seedance 2.0 rolled out in China in early 2026, but its global API on fal.ai went live on April 9, 2026, which is when most Western teams first got hands-on access.

Seedance version timeline from 1.0 in June 2025 to 2.5 announced June 2026, with timing and what each release added.
VersionTimingWhat it added
Seedance 1.0 / 1.0 ProJune 2025First release; multi-shot from text and image; topped the video leaderboard at launch
Seedance 1.5 ProDecember 2025Added native audio (“sound and vision in one”)
Seedance 2.0Early 2026 · global API Apr 9Unified audio-video model; up to 12 reference inputs; currently #1 on Artificial Analysis
Seedance 2.0 MiniMid-June 2026Faster, cheaper tier
Seedance 2.5Announced Jun 23 · GA early July30-second clips, 50 reference inputs, local re-draw editing

03The UpgradesThree claims, each aimed at a familiar limitation.

ByteDance led with three upgrades. Each targets a specific, familiar shortcoming of today's AI video tools — clip length, brand consistency, and the cost of re-rolling a whole clip to change one thing.

Length
Native single-pass clip
30s

Generated in one native pass rather than stitched segments — pitched as the longest single-shot of any model. A continuous 30 seconds covers a full Reels, Shorts, or TikTok unit, or a thirty-second ad, in one unbroken camera move.

vs ~5-15s today
Control
Multimodal reference inputs
50refs

Roughly four times Seedance 2.0's 12-input system. Feed character sheets, product shots, brand assets, environment references, and audio cues together — the controllability story behind brand and character consistency.

images + audio + video
Editing
Local re-draw
1frame

Change a background, swap a product, or replace a subject without altering the rest of the frame — keeping the original motion, camera, lighting, and composition. Demonstrated on the e-commerce product-swap problem.

e-commerce demo
Reported, not yet confirmed
Some secondary coverage adds further claims — native 4K for 2.5 specifically (4K was confirmed for Seedance 2.0), an "industry-first" 3D pre-visualization mode, and a sparse diffusion-transformer architecture. These come from a single secondary source rather than ByteDance's own materials, so treat them as reported and pending general availability rather than established specs.

04The ContextThe lab making these claims already wins the blind tests.

There is no published benchmark for Seedance 2.5 — it is in beta, and any 2.5 score circulating now is invented. What is measurable is the current model. On the Seedance versus Sora and Kling comparison, the pattern is consistent: ByteDance's Seedance 2.0 leads the independent Artificial Analysis Video Arena, a blind human-preference leaderboard, for both text-to-video and image-to-video.

Text-to-video Arena Elo · Seedance 2.0 vs rivals

Source: Artificial Analysis Video Arena (blind human-preference Elo), June 2026 · bars indexed for display · figures are for the shipping Seedance 2.0, not 2.5
ByteDance Seedance 2.0 (720p)Text-to-video with audio · #1 on Artificial Analysis
1,219
#1
Kling 3.0 (1080p Pro)Text-to-video with audio
1,105
Google Veo 3.1Text-to-video with audio
1,094
Read the ranking honestly
Seedance 2.0 also tops the image-to-video boards (around Elo 1,195 with audio). But a #1 arena ranking reflects blind human preference on sampled prompts — not a guaranteed win on every brief, and not a 2.5 number. The takeaway for a 2.5 announcement is narrower and fair: the lab making these 30-second claims is the one whose current model already beats Veo and Kling on preference. For the wider field, see our roundup of the best AI video generators of 2026.

05Pricing & AccessNo 2.5 price yet — but 2.0 sets the economics.

ByteDance has not announced Seedance 2.5 pricing while it is in beta. At general availability it is expected on Volcano Engine's Model Ark API and across Dreamina, Doubao, and CapCut. For a sense of the economics, the current Seedance 2.0 pricing — published on the fal.ai API — is the better guide. Verify current rates before budgeting; video-model pricing moves quickly.

Seedance 2.0 pricing on the fal.ai API by tier, per second and normalized per minute of 1080p.
Seedance 2.0 on fal.aiPrice
720p + audio~$0.30 / sec
720p Fast + audio~$0.24 / sec
1080p + audio~$0.68 / sec
Normalized 1080p / minute~$9 / min
The cost gap is real
Artificial Analysis normalizes Seedance 2.0 to roughly $9 per minute of 1080p, against about $24 per minute for Google Veo 3.1 and roughly $20 for Kling 3.0 Pro. For teams weighing an AI video stack, that difference compounds fast across high-volume short-form output — the kind of work our social media marketing services are built around.

06For MarketersWhere 2.5 actually changes the workflow.

Stripped of the spec sheet, the 2.5 upgrades line up neatly with the things that actually slow agencies down when they try to use AI video in production. Sort your pipelines into the buckets below before you plan around it.

Real spots, not loops
Full social units in one take

A native 30 seconds covers a complete Reels, Shorts, or TikTok unit, or a thirty-second ad, in one coherent camera move — instead of stitching 5-second clips that drift shot to shot.

The headline win
On-brand consistency
50 references hold the brand

Feed character sheets, product shots, and palette in one generation so a subject and product stay steady across the clip. This is the top complaint about gen video, and the upgrade aimed at it.

Strong if it holds up
Variants at scale
One master, many cuts

Local re-draw turns one master clip into product, locale, and seasonal versions without re-rolling the whole video — the SKU-variant workflow ByteDance demoed for e-commerce.

Pilot on real SKUs
Rights and claims
Govern likeness and verify

Build consent and licensing into any workflow using real people or brand IP, and treat the 2.5 numbers as vendor claims until independent tests land. Better clips do not replace a sound distribution plan.

Non-negotiable

The practical move is not to wait for general availability before thinking about it. The current Seedance 2.0, Veo, and Kling are already capable enough to pilot a short-form pipeline; 2.5 mainly raises the ceiling on length and consistency. Pair the tooling with a real short-form video strategy so the better clips actually reach an audience.

07Honest CaveatsWhat to keep front of mind before you plan around it.

A new-model announcement is the moment to be most careful with claims. A few things are worth holding onto:

  • It is beta, not shipping. The early-July date is a vendor target; features and timing can change before general availability.
  • No 2.5 benchmark exists. Any Seedance 2.5 score or Elo circulating today is fabricated — the leaderboard data is for Seedance 2.0.
  • 4K is a 2.0 feature. ByteDance confirmed native 4K for Seedance 2.0; whether 2.5 ships 4K at launch is reported, not officially stated.
  • Preference is not capability. A #1 arena ranking reflects blind human preference on sampled prompts, not a guaranteed win on every brief.
  • Access may be staggered. Enterprise beta first, consumer surfaces later; confirm regional and API availability at launch.

08ConclusionReal news, with measured optimism.

The shape of Seedance 2.5, June 2026

A genuine announcement that extends a lead ByteDance already holds — pending the tests that prove it.

Seedance 2.5 is a genuine piece of news rather than a rumor: ByteDance announced it on June 23, 2026 and pointed to an early-July launch. If the headline claims hold up under independent testing — a native 30-second clip, 50 reference inputs, and frame-local editing — it extends a lead the company already holds, because Seedance 2.0 is the current top-ranked video model at a fraction of Google's and Kling's cost.

The right posture for a marketing team is measured optimism. Treat the 2.5 specifics as claims until launch, keep consent and licensing in the workflow, and use the wait to pilot short-form AI video on the tools that already ship. When 2.5 lands, the teams that benefit will be the ones that already know where AI video fits in their content engine.

Build your AI video strategy

AI video is moving fast — the edge is knowing where it fits.

We help businesses evaluate, benchmark, and deploy AI video tools — Seedance, Veo, Kling and more — for short-form social, paid ads, and product video, without betting the brand on an unproven claim.

Free consultationExpert guidanceProduction-ready workflows
What we work on

AI video engagements

  • Model selection — Seedance / Veo / Kling on cost and quality
  • Short-form pipelines for Reels, Shorts, and TikTok
  • Brand and character consistency across AI-generated clips
  • SKU and locale ad variants from a single master
  • Rights, consent, and disclosure guardrails for AI video
FAQ · Seedance 2.5

The questions teams ask about AI video.

Seedance 2.5 is the next version of ByteDance's text- and image-to-video model, announced on June 23, 2026 at the Volcano Engine FORCE conference. ByteDance's headline claims are a single native 30-second clip, up to 50 multimodal reference inputs, and local re-draw editing that changes one element of a frame without altering the rest. As of the announcement it is in global enterprise beta, not generally available.
Related dispatches

Continue exploring AI video.