AI News · industry
AI Video in November 2026: Sora Is Gone, Chinese Models Lead
OpenAI switched off Sora, 30-second single takes became normal, and seven of the ten best-rated video models are made by Chinese labs or built on one. Here is the November map and what it changes for your ad creative.
As of November 4, 2026, the best-rated AI video model is Alibaba's Wan 3.0, and OpenAI's Sora no longer exists. On the Artificial Analysis text-to-video leaderboard, which ranks models with sound by blind human votes, Wan 3.0 sits first at 1,172 Elo. Four of the top ten come straight from Chinese labs (Alibaba, MiniMax and ByteDance), and three more are built on MiniMax's model. OpenAI shut down the Sora app on April 26 and the Sora API on September 24, 2026. Meanwhile, the length of a single generation doubled: Wan 3.0, Seedance 2.5 and Kling 4.0 all render up to 30 seconds in one pass.
In June we wrote that AI video had gained sound and believable physics and become usable for ad work, and we named Sora 2 and Veo as the leaders. Five months later, one of those is gone and the other has slipped to 22nd. If you make product or ad video, this is the update.
What changed in AI video between June and November 2026?
Three things, roughly in order of how much they matter to a business.
Clips got long enough to tell a whole ad. A 30-second spot used to mean stitching six or seven short clips and hoping the product looked the same in each. Now one generation can carry the full arc. ByteDance's Seedance 2.5 (July 31) went from 15 to 30 seconds per pass. Alibaba's Wan 3.0 generates up to 30 seconds at 30fps with dialogue, music and sound effects. Kuaishou's Kling 4.0, launched October 15, also reaches 30 seconds in a single pass.
References replaced prompting as the main control. Instead of describing your product in words, you hand the model pictures and clips of it. Seedance 2.5 accepts up to 30 images, 10 video clips and 10 audio clips at once. Wan 3.0 takes up to 20 references, including documents and web pages. Kling 4.0 combines up to 15 reference assets with up to 10 keyframes. For a brand, this is the difference between "a bottle that looks a bit like ours" and your actual bottle.
The leaderboard turned over. The models that topped the charts in spring have been passed, and a US lab closed its flagship video product outright.
Who leads AI video right now?
Here is the top of the Artificial Analysis board for text-to-video with audio, as we checked it on November 4, 2026. Elo scores come from blind pairwise votes; the "range" column is the rank the model could plausibly hold given the margin of error.
| Rank | Model | Lab (country) | Elo | Rank range | Weights |
|---|---|---|---|---|---|
| 1 | Wan 3.0 | Alibaba (China) | 1,172 | 1–3 | Closed (API) |
| 2 | Utopai X | Utopai Studios (US), built on MiniMax | 1,169 | 1–4 | Closed |
| 3 | MiniMax H3 (768p) | MiniMax (China) | 1,160 | 2–6 | Open |
| 4 | FLUX 3 Max | fal (US), built with Black Forest Labs | 1,156 | 2–7 | Closed |
| 5 | Seedance 2.5 | ByteDance (China) | 1,155 | 3–7 | Closed |
| 8 | FLUX 3 | Black Forest Labs (Germany) | 1,139 | 7–10 | Closed |
| 9 | Seedance 2.0 | ByteDance (China) | 1,136 | 7–11 | Closed |
| 10 | Gemini Omni Flash 1.1 | Google (US) | 1,129 | 9–11 | Closed |
| 22 | Veo 3.1 | Google (US) | 961 | 22–24 | Closed |
Read the ranges before you read the ranks. The top five sit within 17 points of each other and their ranges overlap, so "Wan 3.0 is the best" really means "Wan 3.0 is first in a tight pack." Places 6 and 7 go to two more models built on MiniMax. Kling 4.0 is too new to appear yet; its predecessor, Kling 3.0 Pro, is the board's fixed 1,000-point anchor.
Google is the highest-ranked US lab with its own base model, and it is clearly moving its bets: its Gemini API deprecations page names Gemini Omni 1.1 Flash as the replacement for every Veo model it has retired, including the Veo 3.1 previews scheduled to shut down from October 22, 2026. Its Gemini Omni 1.1 Flash (August 27) ranks tenth, and Google's pitch for it is control and cost rather than raw ranking: scene extension up to 40 seconds in 10-second steps, first-and-last-frame control, 4K upscaling, and 360p drafts that Google says run up to 60% faster at a third of the cost of its standard 720p output.
Can I run a top video model myself?
For the first time, yes, with caveats. MiniMax H3 is open weights and ranks third overall, the best-rated open model on the board by a wide margin. The 33-billion-parameter release on Hugging Face generates 4 to 15 seconds of video with stereo audio and supports dialogue in 11 languages, including Arabic, Russian and Spanish.
The caveats are real. The open release covers the base model at 768p. MiniMax's prompt-understanding layer and its 2K regeneration step stay hosted, and the model runs under the MiniMax H3 Community License, which asks users in the US, EU, UK and South Korea to submit an application form. Running it also takes serious hardware; MiniMax's own example serves it across four GPUs.
On the Artificial Analysis board, the next open option is LTX-2.5 from Lightricks (Israel), whose Fast version ranks 25th. Black Forest Labs has said an open FLUX 3 Dev covering video is coming, per VentureBeat's launch coverage. Until then, the practical open-weights choice for quality is MiniMax H3.
Why did OpenAI shut down Sora?
OpenAI's help center article on the shutdown gives the dates (the web and app experience ended April 26, 2026; the API ended September 24, 2026), along with export steps and a note that leftover credits can be spent on Codex. It does not give a reason. Axios reported the decision on March 24, and NBC News reported that Disney's deal with OpenAI, struck the previous December, would be dissolved.
We won't guess at the internal reasons. The lesson for anyone building a content workflow is simpler. A model we named a category leader in June is now gone, with about six months between the announcement and the API switching off. If your product video pipeline was wired to Sora, you rebuilt it this autumn. We made the general case in why you shouldn't build your business on one AI model; Sora is the cleanest example yet.
What does this mean for your ads?
Advertisers were already moving. In the 2026 IAB Digital Video Ad Spend & Strategy report, nearly two in three video ad buyers said they use generative AI for video creative, up from half in 2025. IAB expects a third of their ad assets to use generative AI this year (up from a quarter) and 43% by 2027. Among smaller buyers, 96% said they were not satisfied with their current level of AI use in creative.
So if you run a clinic or a local store and have not started, you are not late, but the gap is opening. Here is how the November models map to real jobs:
- A full 15 to 30 second product spot from one generation. Wan 3.0, Seedance 2.5 or Kling 4.0, fed with your real product photos as references.
- Lots of cheap drafts before you commit. Gemini Omni 1.1 Flash's 360p draft mode exists for exactly this.
- Your own footage, re-staged. Seedance 2.5 can swap a green-screen background or change the camera move on a clip you already shot, while keeping the subject.
- Keeping everything in-house. MiniMax H3, if you have the GPUs and accept the license terms.
One thing has not changed since June: a better-looking clip is not the same as a better-performing ad. More 30-second takes means more chances to publish something off-brand or non-compliant. Health and beauty businesses should take extra care here, because a single unsupported before-and-after claim can get an ad rejected.
Where does Entagl fit?
Entagl doesn't build video models. On accounts with Studio turned on, it can turn a product from your catalog, including items imported from Shopify, into multi-scene product video. Then, when you approve an asset for ads, the Ads Co-Pilot (rolling out by account) reviews it: it checks for a hook in the first seconds, lists the claims the ad makes, flags likely problems with Meta's ad policies (plus sector rules for supported industries such as dental and aesthetics), and gives an ad-readiness score. By default it proposes and you approve.
That review step matters more as models produce more. If you are working out how to turn volume into results, our guide to scaling ad creative with AI covers the testing side, and AI avatars and digital humans in 2026 covers the talking-presenter formats.
What still doesn't work?
Be skeptical of any demo reel, including the labs' own. ByteDance's Seedance 2.5 announcement says there is "still room for improvement" in the physics of complex motion and in scenes where several people interact. Thirty seconds also means thirty seconds for a hand or a label to drift. And leaderboard Elo measures what voters prefer in short side-by-side tests, not whether a clip sells your product. Test your own prompts on two or three models before you choose one.
FAQ
What is the best AI video model in November 2026?
On the Artificial Analysis text-to-video leaderboard (with audio), checked November 4, 2026, Alibaba's Wan 3.0 ranks first at 1,172 Elo, followed closely by Utopai X, MiniMax H3, fal's FLUX 3 Max (built with Black Forest Labs) and ByteDance's Seedance 2.5. Their ranges overlap, so treat the top five as a near-tie and pick by job, cost and controls.
Is OpenAI Sora still available?
No. According to OpenAI's help center, the Sora web and app experiences were discontinued on April 26, 2026, and the Sora API on September 24, 2026. Users were told to export their content.
How long can AI-generated video clips be now?
Wan 3.0, Seedance 2.5 and Kling 4.0 each generate up to 30 seconds in a single pass. Google's Gemini Omni 1.1 Flash can extend a scene in 10-second steps to 40 seconds total, and Seedance 2.5 supports multiple rounds of extension.
Is there a good open-source AI video model?
MiniMax H3 is the highest-rated open-weights model on the Artificial Analysis board, ranking third overall as of November 4, 2026. It is released under the MiniMax H3 Community License, and the open release covers 768p generation; the 2K step remains a hosted service. On the same board, Lightricks' LTX-2.5 is the next open option.
Can small businesses use these models for real ads?
Yes, and many already do: IAB's 2026 report found nearly two in three video ad buyers use generative AI for video creative. The safe way is to use your real product images as references, generate several versions, and review each one for brand accuracy and ad-policy compliance before spending on it.
Making product video is the easy part now. Knowing which clip to run is not. See how Entagl's Studio and Ads Co-Pilot help turn your catalog into reviewed, ready-to-test video. Book a 30-minute demo.
Sources: Artificial Analysis, AA-Video-T2V v2.1 leaderboard (checked November 4, 2026); OpenAI Help Center, Sora discontinuation; Axios; NBC News; ByteDance Seed, Seedance 2.5; Alibaba Cloud, Wan 3.0 guide; Kuaishou, Kling 4.0 launch; Google, Gemini Omni 1.1 Flash; MiniMax H3 on Hugging Face; Lightricks LTX-2.5 on Hugging Face; VentureBeat, FLUX 3 launch; IAB, 2026 Digital Video Ad Spend & Strategy report; Google, Gemini API deprecations. Model details reflect vendor announcements and leaderboard standings as of November 4, 2026.