AI Video Models in 2026: Sora Out, China Leads

The best AI video model in the world right now is Chinese, and the most famous one is being switched off. As of July 2026, ByteDance’s Dreamina Seedance 2.0 sits at the top of Artificial Analysis’s blind-vote Video Arena, with Alibaba’s HappyHorse models, Kuaishou’s Kling 3.0 and Alibaba’s open-weight Wan 2.7 filling most of the seats behind it. Google’s Veo 3.1 is the highest-placed Western model. OpenAI’s Sora, the product that made “AI video” a household phrase in 2024, does not appear at all: OpenAI is discontinuing it, having already shut the app and with the API scheduled to go dark on September 24, 2026.

That is the whole 2026 story in one paragraph: the frontier moved to China, the price of frontier-grade video collapsed, and the marquee American launch turned into a case study in how expensive video generation is to actually run. Here is the evidence.

Who leads right now

Artificial Analysis runs a video version of the same blind-comparison Elo arena it uses for language models: users are shown two clips generated from an identical prompt, they pick the better one, and votes roll up into an Elo score. It is a preference ranking, not ground truth, and it splits the board by whether generated audio is included, because native sound is now a real differentiator. These are the top text-to-video models with audio, read on July 4, 2026:

The top of the board is almost entirely Chinese. Veo 3.1 is close behind on raw preference and remains the strongest Western entry, but the six seats above it belong to four Chinese companies. The image-to-video board tells the same story: Seedance 2.0 leads there too, with xAI’s grok-imagine preview the only non-Chinese model in the top handful. Strip audio out of the comparison and Alibaba’s HappyHorse-1.0 takes first place outright.

Preference scores only matter to a business if the winning models are also affordable and available. On both counts, 2026 broke in the same direction.

Quality is no longer the expensive part

For most of the short history of generative video, the assumption was that the best model would also be the most expensive to run. In July 2026 that link is broken. The models topping the arena are among the cheapest to call, and the priciest option on the board is the Western flagship.

The price gap is the same shape as the one we documented for language models in why Chinese AI models are so cheap, and it has the same causes: efficient training, cheap domestic power and an open-weight strategy that prices for adoption rather than margin. Video just makes it more visible, because a second of generated video is a far heavier compute unit than a token of text.

Two caveats keep this honest. First, video pricing is genuinely messy: rates depend on resolution, on speed tier, on whether native audio is switched on, and on whether you buy from the model maker or a reseller, so the numbers above are representative list rates rather than a single canonical price. Kling, for example, meters access in credits rather than dollars, and Chinese providers rarely publish a clean per-second dollar rate. Second, Wan 2.7 is open weight under an Apache 2.0 license, so if you own the hardware the marginal cost is your own electricity, not a per-second fee at all. The cloud rate shown is what a hosted API charges to run it for you. You can check the current per-token equivalents for the underlying model families on our AI model price tracker.

The 2026 release wave

The current standings are the product of a compressed six-month sprint. Almost every model on the board shipped or shipped a major version this year.

Runway rounds out the field. Its Gen-4.5, released at the end of 2025, is not a leaderboard chart-topper but remains the professional workflow favorite for the same reason it always has been: reference-image character consistency, motion brush, camera control and a real editor around the model. On Runway’s own pricing, a second of Gen-4.5 costs 12 credits, which works out to roughly $0.10 to $0.23 per second depending on which monthly plan you spread the credits across.

Why Sora is being switched off

The most instructive event of 2026 is not a launch. It is a shutdown. OpenAI announced in March that it would discontinue Sora: the web and app experiences went offline on April 26, 2026, and the API is scheduled to follow on September 24, 2026.

The reason, as reported across the technology press, was economics rather than capability. Sora was reportedly costing on the order of a million dollars a day to run against a small fraction of that in revenue, its download numbers had fallen sharply from a late-2025 peak, and a planned licensing deal with Disney for character rights had collapsed. Treat the specific figures as reported rather than confirmed, because OpenAI has not published a full accounting. The direction is the point: a frontier US lab decided that operating a consumer video model at scale was not worth the compute bill, in the same quarter that Chinese labs were giving comparable quality away at a few cents a second.

What to actually use

The leaderboard is not a buying guide. What you should reach for depends on the job.

• Best all-around quality with sound: Dreamina Seedance 2.0 leads the arena and generates coherent audio alongside the picture. Kling 3.0 is a close alternative with strong multi-shot control and a cheaper entry price.
• Best if you live in the Google or enterprise stack: Veo 3.1 is the strongest Western model, with native 48kHz audio and up to 4K output. Veo 3.1 Lite covers high-volume, cost-sensitive jobs at roughly a tenth of the flagship price.
• Best professional editing workflow: Runway Gen-4.5, for reference-driven character consistency, camera control and the surrounding editor, when creative control matters more than a top arena score.
• Best if you want to own the model: Wan 2.7. Open weights under Apache 2.0 mean you can self-host, fine-tune and run it without a per-second bill, and it still ranks in the arena’s top six.
• If you were building on Sora: migrate now. The API is gone at the end of September 2026, and Veo, Kling, Seedance and Runway all offer a path off it.

For how these makers stack up beyond video, see our AI model leaderboard and the AI benchmarks directory, and for the broader cost argument, why Chinese AI models are so cheap.

Bottom line

The state of AI video in July 2026 is a two-part inversion. Quality leadership has moved to Chinese labs, with Seedance, HappyHorse, Kling and Wan holding the top of the independent arena and Google’s Veo 3.1 the best Western answer. And the cost of frontier-grade video has fallen to a few cents a second, cheap enough that the economics, not the model, now decide who ships. Sora’s shutdown is the clearest signal of the new rules: in the most compute-hungry corner of consumer AI, the cheapest capable model wins, and right now the cheapest capable models are not American.

Sources

• Artificial Analysis (2026). Text-to-Video Arena Leaderboard and Image-to-Video Arena Leaderboard. Artificial Analysis. Read July 4, 2026. https://artificialanalysis.ai/video/leaderboard/text-to-video
• OpenAI (2026). What to know about the Sora discontinuation. OpenAI Help Center (documentation). https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation
• Google (2026). Build with Veo 3.1 Lite, our most cost-effective video generation model. The Keyword, Google. March 31, 2026. https://blog.google/innovation-and-ai/technology/ai/veo-3-1-lite/
• Google Cloud (2026). Vertex AI generative AI pricing (Veo section, documentation). https://cloud.google.com/vertex-ai/generative-ai/pricing
• Runway (2026). Pricing (documentation). https://runwayml.com/pricing
• KlingAI (2026). API Service pricing (documentation). https://kling.ai/dev/pricing
• CNBC (2026). Alibaba, ByteDance and Kuaishou among new China AI model releases. CNBC (secondary, as reported). February 14, 2026. https://www.cnbc.com/2026/02/14/new-china-ai-models-alibaba-bytedance-seedance-kuaishou-kling.html