After Sora's Fall, 'HappyHorse' Takes the Crown: The AI Video Race Reshuffles
Sora died faster than I expected.
Last month, OpenAI announced they were shutting down the Sora application. Less than a year since its release. I still remember when that demo video broke the internet—who would have thought it would become history so quickly?
But what’s more surprising is what replaced it.
In early April, an anonymous model called HappyHorse-1.0 quietly appeared on Artificial Analysis’s Video Arena blind test leaderboard—no company name attached. It topped the image-to-video chart with 1411 points, leaving ByteDance’s Seedance 2.0 in the dust by about 55 points. First place in text-to-video too, with 1379 points.
Honestly, this plot twist is pretty dramatic.
Sora’s death, in my view, wasn’t a technology problem—it was a business model problem. Video generation is incredibly compute-intensive. Industry estimates suggest generating one minute of video can cost tens of dollars. And users’ willingness to pay? Way below that number.
This creates an awkward situation: the better you do, the more you lose. OpenAI ran the numbers and realized burning cash on this made no sense. Better to kill it.
But HappyHorse’s rise proves something: the market demand exists—Sora’s business model just couldn’t capture it.
As for HappyHorse’s identity, theories abound. Some speculate it’s an Alibaba model (given their recent moves in AI video), others suggest it’s from an overseas team working anonymously. I think the identity doesn’t matter—what matters is that it proves you can succeed without the OpenAI brand halo, as long as you have real capability.
This reminds me of Runway’s Gen-4. Released this April, it supports 8K resolution and cinema-grade RAW output—highly regarded in professional circles. Runway’s approach differs from Sora’s; they’re more focused on B2B and professional creators rather than trying to build a consumer product for everyone.
Looking back, that strategy seems correct. Consumer users want “free and good,” but video generation’s cost structure makes that unsustainable. B2B clients, on the other hand, are willing to pay premium prices for quality output—a more viable business model.
Jimeng AI’s recent price hike tells the same story. Two-minute video production costs jumped from 5 yuan to 40 yuan—an 8x increase. Users are understandably unhappy, but from a business perspective, this might be a necessary adjustment.
Watching the AI video race lately, one feeling keeps growing stronger: we’re transitioning from a “tech demo” phase to a “business落地” phase.
Tech demo phase: whoever has the most impressive effects wins. Business落地 phase: whoever has the healthiest business model survives longest.
Sora represented the former; Runway and possibly HappyHorse represent the latter.
Another trend worth watching: the rise of Chinese models. Alibaba, ByteDance, and Kuaishou are all investing heavily. Kling AI, Jimeng AI, Tongyi Wanxiang—each has its strengths. While generation quality may still lag top overseas models, cost control is better, and they understand domestic user needs more deeply.
I tried Alibaba’s Tongyi Wanxiang the other day. Generation speed is impressive, and Chinese comprehension is better than overseas models. For short video content creation, it’s more than sufficient.
So my assessment: the “Sora Era” of AI video is over. We’re entering the “Hundred Flowers Era.” No more single dominant player—multiple competitors will develop in different market segments. Professionals use Runway, casual users use domestic tools, and anonymous talents might surprise us anytime.
This is interesting because it shows there are no permanent kings in AI. Even OpenAI can stumble in specific domains. For entrepreneurs, this means opportunity—the blind spots of giants are breakthrough points for startups.
One final question: how often do you use AI video tools? What do you mainly use them for?