AI video generation just moved to China — and Western media barely noticed

On April 26, 2026, OpenAI did something nobody saw coming: it shut down Sora. Not the API — that lasts until September 24. The consumer app and web experience went dark, and every account holder got the same message: export your videos by the deadline or they're gone forever. The shutdown was the cleanest possible signal that creative AI is no longer where OpenAI is placing its bets. Compute is going to coding. Compute is going to enterprise. Compute is not going to the world's most expensive video toy.

Then something unexpected happened to the leaderboard.

For roughly 60 days starting in early February, xAI's Grok Imagine 1.0 sat at #1 on Artificial Analysis — the only blind, human-judged AI video arena that real creators trust. It beat Runway Gen-4.5. It beat Kling 2.5 Turbo. And it beat Google's Veo 3.1 at one-seventh the price. Grok Imagine charged $4.20 per minute with audio. Veo charged $12. Sora 2 Pro charged $30. The Batch from DeepLearning.AI called the spread "a seven-fold price gap with room to fall." For the first time since the field existed, the cheapest model was also the best.

Then the field moved.

As of May 2026, the Artificial Analysis Image-to-Video leaderboard reads: HappyHorse-1.0 at #1 (1,397 Elo). ByteDance's Dreamina Seedance 2.0 at #2 (1,347). Grok Imagine at #3 (1,327). Veo 3.1 fell out of the top three. The Text-to-Video leaderboard tells the same story: HappyHorse-1.0, Dreamina Seedance 2.0, Kuaishou's Kling 3.0 in the top three; Grok Imagine slipped to #5; Veo 3.1 hovers at #4. Read those lists again. Three of the top five models on the most respected blind benchmark in this space are now from Chinese labs. None of those Chinese models existed eighteen months ago.

Why this matters: the Western lead in consumer AI video lasted three months. That is not a moat. That is a head-start that evaporated. ByteDance ships Dreamina inside CapCut, which already has 800 million monthly active users. Kuaishou ships Kling inside its short-video super-app, which has 600 million daily users. Distribution is not a concern for them. Pricing is not a concern for them. The only frontier-lab moves that matter now are physics fidelity, multi-shot consistency, and 4K output — and even on those, Veo 3.1 is the lone Western model still competitive.

The numbers nobody is repeating: Grok Imagine generates a 15-second video with audio in 17 seconds. Veo 3.1 takes longer at higher cost. Sora 2 Pro is dead at any speed or price. Kling 3.0 ships natively at 1080p. Dreamina Seedance 2.0 generates 720p with audio inside CapCut for free for ByteDance subscribers. The model OpenAI killed because it was too expensive to run is now being out-shipped by Chinese platforms that treat video generation as a free feature on top of products users already pay for.

The countervailing angle every American AI investor will want to read twice: this is not a "China is catching up" story. It is a "China already passed and the markets did not reprice" story. The moats that Sora's launch implied — proprietary data, top-tier compute, and the consumer brand of OpenAI — turned out to last 90 days. If the same trajectory holds for code generation, for agentic systems, for whatever ships next, the Western frontier-lab thesis needs an honest stress test. The talent argument doesn't survive contact with HappyHorse-1.0 sitting at #1 on a blind leaderboard against models that cost ten times more to run.

What to watch over the next 1-3 months: (1) whether HappyHorse-1.0 ends up being a known Chinese lab in disguise or a genuine new entrant — the team is currently anonymous, which matters. (2) Google's response with Veo 3.5 or Veo 4, and whether it can recapture the top-three. (3) The September 24 Sora API shutdown — that's when every enterprise customer locked into Sora has to decide between Veo, Grok, or moving to a Chinese provider. (4) Whether the US Bureau of Industry and Security intervenes if a major American media customer moves to ByteDance or Kuaishou for video gen.

The takeaway: the consumer AI video lead held by American labs lasted exactly one season. Sora launched in February 2024. Grok Imagine took #1 in February 2026. By May 2026 the top three are Chinese. The race is over for Q1 — and the winners are not the names anyone outside of WaveSpeed and Artificial Analysis is repeating.