OpenAI pulled the plug on Sora this week. The app goes dark in April; the API follows in September. Everyone's writing eulogies. I want to write something different: a thank-you note. Because Sora dying is the best thing that's happened to AI video in over a year.
The numbers that killed it
Sora cost OpenAI roughly 15 million per day to operate. Fifteen. Million. Dollars. Per day. Against that, the app generated 2.1 million in lifetime App Store revenue. Not per month — total, across its entire existence. Bill Peebles, OpenAI's own head of Sora, admitted publicly that "the economics are completely unsustainable."
Downloads peaked at 3.33 million in November 2025 — the hype month — then cratered 66% to 1.13 million by February. The Disney partnership, reportedly worth $1 billion in licensing and investment, collapsed over deepfake liability concerns before any money changed hands.
This wasn't a product that failed to find product-market fit. This was a research demo that got shoved into a consumer app because the demo reel at DevDay 2024 went viral and someone in a boardroom thought virality meant demand.
The demo-to-product gap nobody talks about
Here's the uncomfortable truth about Sora that the AI video community danced around for sixteen months: the best Sora outputs were cherry-picked. Everyone in production knew this. You'd queue up a generation, wait minutes, get something unusable, try again, wait again, maybe on the fourth or fifth attempt get something close to what the marketing materials promised.
Runway figured this out early. So did Kling. The competitive advantage in AI video isn't peak quality — it's floor quality. A tool that gives you a reliable 8 out of 10 every time beats one that occasionally hits 10 but usually delivers a 4. Sora optimized for ceiling. The market wanted floor.
This is why OpenAI built Sora as a social platform instead of a professional tool. They couldn't sell it to studios — the consistency wasn't there, and the deepfake liability made legal departments break out in hives. So they tried the consumer play: make it TikTok but with AI. Except nobody needed TikTok-but-with-AI when TikTok-with-CapCut already existed and was free.
The $15M/day question
Why was it so expensive? Video generation demands roughly 1,000x the compute of equivalent text generation. That's not an engineering problem you optimize away with a better architecture — it's a structural reality of the medium. Every frame is an image. Every second is dozens of frames. Temporal coherence across those frames requires the model to maintain spatial state across a sequence that grows linearly with duration.
OpenAI threw their most powerful hardware at it and still couldn't get the unit economics below catastrophic. That tells you something fundamental about where we are in the cost curve for AI video: the models work, but they don't work cheaply, and no amount of Sam Altman's fundraising changes thermodynamics.
What actually works right now
The market Sora leaves behind is healthier than the one it entered. Instead of one overhyped monolith, we have genuine specialization:
Runway Gen-4 owns the professional tier. Best temporal consistency, best motion control, 4K output, 15-second clips. If you're cutting a commercial or need footage that an editor can actually work with in Premiere, this is the tool. $15–76/month depending on volume. The tradeoff: it's expensive per second and conservative in what it'll generate.
Kling 3.0 dominates the volume tier. Kuaishou priced it at roughly $0.15 per second of video — about 40% cheaper than Runway for comparable quality. Human physics and facial expressions are its strength. If you're a social media team producing thirty clips a week, Kling is where you end up. 1080p cap, 30-second max duration.
Veo 3 from Google is the quiet powerhouse. 4K, 60-second clips, excellent prompt adherence for cinematic content. It's buried inside Google's ecosystem, which limits adoption, but the raw output quality arguably leads the market right now.
Seedance 1.5 Pro is the open-source dark horse at $0.044 per second — roughly a third of Kling's price. Quality is a tier below the leaders, but for storyboarding, prototyping, and internal content, the economics are hard to argue with.
This segmentation is good. It means creators can match tools to workflows instead of wrestling with a single platform that tries to be everything and nails nothing.
The real lesson
Sora's failure wasn't about technology. The model was genuinely impressive — anyone who used it knows that. The failure was strategic. OpenAI treated video generation as a product vertical when it should have been infrastructure. They built a walled-garden app when creators wanted an API and an export button. They chased viral consumer adoption when the actual paying customers were production studios who needed reliability guarantees and IP indemnification.
The team isn't disbanded — they're pivoting to "world simulation" for robotics, which is probably where this technology belonged all along. Training a model to understand how physical objects move and interact is more valuable inside a robotics stack than inside a content creation app that hemorrhages money.
Meanwhile, Runway keeps shipping. Kling keeps undercutting. Veo keeps quietly improving. The tools that survive in synthetic video are the ones built by teams that understood from day one that the job isn't generating impressive demos — it's fitting into someone's Tuesday afternoon editing session.
Sora gave us the most spectacular AI video demo the world had ever seen. It also proved, definitively, that a spectacular demo and a viable product are completely different things. The next time someone shows you a breathtaking AI generation and says "this changes everything," remember the $15 million per day.
It doesn't change anything until it ships.