Go To AgencySeedance 2.5: ByteDance broke the 30-second barrier for AI video, and where it ranks now
On 23 June 2026, ByteDance announced Seedance 2.5: 30-second AI video clips in one continuous generation, with up to 50 reference inputs. Released in July, it now ranks third on the Artificial Analysis text-to-video arena with audio, second on image-to-video. What was announced, the Seedance lineage from the papers, the leaderboard on 17 October 2026, and the catches.
In short
- ByteDance announced Seedance 2.5 on 23 June 2026 at Volcano Engine's FORCE conference: clips of up to 30 seconds in one generation, with no stitching, and up to 50 reference inputs; it was released in July.
- On 17 October 2026, Artificial Analysis ranks it third on text-to-video with audio (Elo 1160, behind Wan 3.0 and Utopai X) and second on image-to-video with audio (1138).
- The lineage, from ByteDance's papers: Seedance 1.0 (June 2025), 1.5 pro with native joint audio-video (December 2025), 2.0 (February 2026, 4 to 15-second clips).
- Catches: blind-vote rankings are not your use case, no official per-second price we could verify, and a data governance question for European companies.
On 23 June 2026, at Volcano Engine's FORCE conference, ByteDance announced Seedance 2.5, the next generation of its AI video model, and one number travelled faster than the rest: 30 seconds. Not 30 seconds stitched from short clips: one continuous generation, scene changes and tempo shifts included, according to The Decoder. In a field where most models hand you a few seconds at a time, that resets expectations.
This article was first published on announcement day and revised on 17 October 2026. Seedance 2.5 has since been released, and it now appears on the independent Artificial Analysis Video Arena, which it did not at the time. The leaderboard figures from the first version, which we could no longer trace to a dated source, have been replaced by the board as it stands on 17 October 2026.
What ByteDance announced
| Seedance 2.5 | Detail | Source |
|---|---|---|
| Announced | 23 June 2026, at Volcano Engine's FORCE conference; Volcano Engine is ByteDance's cloud business | The Decoder |
| Headline capability | Single clips of up to 30 seconds, with no stitching | The Decoder |
| References | Up to 50 additional inputs, such as reference images and audio, for scenes with several characters | The Decoder |
| Editing | A generated video can be edited afterwards while keeping its visual style | The Decoder |
| Release | Announced for early July; listed by Artificial Analysis as released in July 2026 | The Decoder; Artificial Analysis |
At the same conference, ByteDance said Seedance 2.0 now supports native 4K with 10-bit colour, and presented the Doubao 2.1 Pro language model, the Seedream 5.0 Pro image model and the Seed-Audio 1.0 audio model.
Where it ranks: the independent leaderboard
The Artificial Analysis Video Arena ranks models by blind human preference. On 17 October 2026, on the text-to-video board with audio:
| Rank | Model | Creator | Elo | Released |
|---|---|---|---|---|
| 1 | Wan 3.0 | Alibaba | 1174 | August 2026 |
| 2 | Utopai X | Utopai Studios | 1169 | September 2026 |
| 3 | Dreamina Seedance 2.5 | ByteDance Seed | 1160 | July 2026 |
| 4 | MiniMax H3 (768p) | MiniMax | 1159 | July 2026 |
| 5 | MiniMax H3 Max | MiniMax (via Fal) | 1155 | August 2026 |
| 6 | FLUX 3 | Black Forest Labs | 1146 | July 2026 |
| 7 | Dreamina Seedance 2.0 | ByteDance Seed | 1136 | March 2026 |
| 8 | Gemini Omni Flash 1.1 | 1132 | August 2026 |
On the image-to-video board with audio, Seedance 2.5 is second, at 1138, two points behind MiniMax H3 (768p) at 1140; Artificial Analysis notes that their confidence intervals overlap, so the gap is not meaningful. Two cautions when reading these boards: the no-audio views give different orders, and Elo scores move as new models and votes arrive, so compare positions on the same date and the same view.
The takeaway: Seedance 2.5 sits in the top tier, clearly ahead of its predecessor, but no longer alone at the top. In four months, Alibaba, MiniMax, Black Forest Labs and Google all shipped models in the same range.
The Seedance line, in dates
- Seedance 1.0 (June 2025). Tech report on arXiv (10 June 2025): native multi-shot generation, text-to-video and image-to-video learned jointly, and a 5-second 1080p video generated in 41.4 seconds on an NVIDIA L20.
- Seedance 1.5 pro (December 2025). Native, joint audio-video generation with a dual-branch Diffusion Transformer, and multilingual lip-sync (arXiv, 15 December 2025).
- Seedance 2.0 (February 2026). Officially released in China in early February 2026; clips of 4 to 15 seconds at 480p and 720p, with up to 3 video clips, 9 images and 3 audio clips as references on its open platform (arXiv, 15 April 2026).
- Seedance 2.5 (announced 23 June, released July 2026). Up to 30 seconds in one pass, up to 50 inputs.
Under the hood
The Seedance family is a diffusion-transformer lineage, and here the primary sources are solid. Seedance 1.0's report describes an architecture that natively supports multi-shot generation and learns text-to-video and image-to-video together, with a tenfold inference speed-up from distillation. Seedance 1.5 pro moved to a dual-branch Diffusion Transformer with a cross-modal joint module, which generates picture and sound together instead of dubbing sound afterwards. Seedance 2.0 adopts a unified architecture for joint audio-video generation that accepts four input modalities: text, image, audio and video. ByteDance has not published a technical report specific to 2.5, so its internals are not documented.
The catches to price in
- Ranking is not a guarantee for your use. A blind-vote Elo measures preference on short prompts. A 30-second brand film with a recurring character is a different test: run it on your own briefs before committing.
- Pricing: check the official consoles. Seedance is sold through Volcano Engine (China) and BytePlus ModelArk (international). We did not find an official per-second price we could verify; the figures circulating are reseller estimates.
- Data governance. Volcano Engine and BytePlus belong to ByteDance. For a European company handling client footage or personal data, where generations are processed and under which law is a question to settle before sending anything sensitive.
- Likeness and provenance. A model that holds a face for 30 seconds is also a better tool for impersonation. Check how outputs are marked before publishing generated people.
Our read: what it changes in a video workflow
What follows is our analysis.
The 30-second single pass is not a gimmick. Much real-world video, a product explainer, a social ad, an opening sequence, lives between 15 and 30 seconds, and stitching short AI clips is exactly where consistency breaks: a face drifts, the lighting jumps, the motion stutters at the cut. A model that holds one generation for 30 seconds, with dozens of references to lock character and style, attacks that problem head-on.
The practical advice is the same as for every model: treat it as a replaceable component in your process, not as the process. The October leaderboard proves the point: the top spot changed hands several times in four months. Keep a second provider wired up, judge on your own briefs, and settle the data question before any sensitive footage goes near an API. For the rest of that summer's AI wave, read our pieces on Meta Muse Image, Figma Motion and GLM-5.2.
Frequently asked questions
What is Seedance 2.5?
ByteDance's AI video model announced on 23 June 2026 at the FORCE conference of Volcano Engine, its cloud business. It generates single clips of up to 30 seconds without stitching, accepts up to 50 additional inputs such as reference images and audio, and lets you edit a generated video while keeping its style. Artificial Analysis lists it as released in July 2026.
Is Seedance 2.5 the best AI video model?
It is among the best, not alone at the top. On 17 October 2026, Artificial Analysis ranks it third on its text-to-video arena with audio (Elo 1160), behind Alibaba's Wan 3.0 (1174) and Utopai X (1169), and second on image-to-video with audio (1138), statistically level with MiniMax H3.
How long a video can Seedance 2.5 generate?
Up to 30 seconds in a single clip, one continuous generation with scene and tempo changes, according to ByteDance's announcement. Its predecessor Seedance 2.0 generates clips of 4 to 15 seconds, according to its technical report.
How much does Seedance 2.5 cost?
We did not find an official per-second price we could verify. Seedance is sold through Volcano Engine in China and BytePlus ModelArk internationally; check their consoles, since the figures circulating online are reseller estimates.
What is the difference between Seedance 2.0 and 2.5?
Seedance 2.0, released in China in early February 2026, generates 4 to 15-second clips at 480p and 720p with up to 3 video clips, 9 images and 3 audio clips as references. Seedance 2.5 goes up to 30 seconds and 50 inputs, and ranks higher: 1160 against 1136 on the Artificial Analysis text-to-video arena with audio on 17 October 2026.
Can a European company use Seedance with client footage?
Only after settling data governance: Volcano Engine and BytePlus are ByteDance services, so check where generations are processed and under which law before sending client or personal data. Keep a second provider available so a single API never holds your production hostage.
Read next
- July 8, 2026Grok 4.5: xAI's fast, cheap frontier model, its benchmarks and what replaced it
- July 7, 2026Meta's Muse Image: an AI that generates, edits, codes and second-guesses itself, and a privacy U-turn
- July 1, 2026Anthropic switched Claude Fable 5 back on: inside the 19-day suspension and the redeployment
Is your project decided?
Present it in writing. We read everything, then we answer.
Pitch a project