Sora Is Dead: Why Creative AI Loses to Productive AI
OpenAI shut down Sora after 6 months. $15M/day in costs, $2.1M total revenue. The economic rift between creative and productive AI, explained.
Fifteen million dollars a day. That’s what Sora was costing OpenAI in inference compute, according to estimates from Forbes and Cantor Fitzgerald analyst Deepak Mathivanan. On the other side? $2.1 million in total revenue over its entire lifetime. Six months of operation, and the cost-to-revenue ratio is staggering: every single day of running Sora burned more money than the app had ever earned.
On March 24, 2026, OpenAI pulled the plug. No bug, no pivot — a clean kill. And in the aftermath, a billion-dollar partnership with Disney evaporated in under an hour.
But this isn’t just the story of a failed product. It reveals a fundamental economic rift in the AI industry: creative tools burn cash, productive tools generate it. And this rift is about to reshape the entire landscape.
The Numbers Behind Sora’s Death
Let’s start with the raw facts, because they speak for themselves.
Sora 2 launched on September 30, 2025, with a standalone iOS app and a consumer-facing experience. By November, downloads peaked at 6.1 million. Then the slide began: 1.1 million in February 2026 — a 66% drop. Monthly active users fell below 500,000.
On the revenue side, the numbers come from Appfigures, the mobile intelligence firm: $2.1 million in total lifetime revenue through in-app purchases. That’s it. For context: that’s less than what Sora cost to run for a single day.
According to Cantor Fitzgerald analyst Deepak Mathivanan (reported by CNBC), each 10-second video clip generated by Sora cost OpenAI roughly $1.30 in compute. With millions of users testing the tool — often out of curiosity rather than genuine need — the bill spiraled out of control.
Bill Peebles, Sora’s lead at OpenAI, publicly admitted that “the economics are completely unsustainable.”
To grasp the scale of the financial hole: OpenAI was generating approximately $25 billion in annualized revenue in early 2026, according to Sacra. Sora alone was burning the equivalent of $5.4 billion per year. A single product was absorbing more than 20% of the company’s revenue — for virtually zero return.
Why AI Video Is So Expensive
Sora’s problem wasn’t that it didn’t work — it’s that it worked too well to be profitable.
Generating one minute of video with a diffusion model like Sora consumes 10 to 15 times more compute than an equivalent ChatGPT conversation, according to estimates from tech-insider.org. It’s the very nature of the task: where an LLM generates text token by token — a relatively lightweight process — a video model must synthesize hundreds of coherent frames, maintain physics-accurate motion, manage lighting, and do it all at high resolution.
Sora’s diffusion transformer architecture was never designed for inference efficiency. It was built as a research showcase — to demonstrate what was possible, not to run at scale with positive margins. When OpenAI tried to turn it into a consumer product, the architecture couldn’t keep up.
Competitors took the opposite approach. Runway with Gen-4 Turbo charges $0.05 per second, roughly $0.50 for a 10-second clip — 2.5 times cheaper than what Sora cost OpenAI internally. Kling 2.5 by Kuaishou pushed optimization even further, producing a 10-second clip in 45 to 75 seconds, compared to 3 to 8 minutes for Sora.
The lesson: these competitors treated inference cost as a first-order engineering constraint, not a problem to solve after launch. Sora did the opposite — and paid the price.
Creative vs. Productive: The Real Signal
Here’s the core thesis of this article, and it goes far beyond Sora.
The AI industry is splitting into two distinct economies with radically different dynamics:
| Productive AI | Creative AI | |
|---|---|---|
| Examples | Claude Code, GitHub Copilot, AI agents, ChatGPT Enterprise | Sora, Midjourney, Runway, Kling |
| Revenue model | Recurring subscriptions, enterprise licenses | Freemium, one-time purchases, credits |
| Cost per use | Low (text tokens) | High (GPU-intensive) |
| Retention | Strong (embedded in daily workflow) | Weak (sporadic use, curiosity-driven) |
| Willingness to pay | High (measurable ROI: time saved) | Low (perceived as a novelty) |
| Usage cycle | Daily, professional | Sporadic, recreational |
The fundamental difference boils down to one thing: measurable ROI.
When a developer uses Claude Code or Copilot, they know exactly how much time they’re saving. A company deploying ChatGPT Enterprise can quantify the productivity gains. That direct ROI justifies a monthly subscription of $20 to $200 — and it recurs.
When a user generates a video of a cat surfing a wave with Sora? The thrill is instant, the shares go viral… and there’s zero reason to come back tomorrow. Sora’s 30-day retention had dropped below 10%. People downloaded it, played around, and left.
This is exactly what Fidji Simo, OpenAI’s CEO of Applications, described when she talked about stopping the “side quests.” Sora had become the optional boss fight in a game where the real objective is enterprise — and where AI agents are reshaping the competitive landscape.
The Disney Debacle: When $1 Billion Vanishes in 30 Minutes
The Disney episode deserves a closer look, because it perfectly illustrates the systemic risk of AI partnerships.
In December 2025, Disney announced a $1 billion investment in a partnership with OpenAI around Sora. The deal was supposed to let users generate videos featuring over 200 characters from Disney, Marvel, Pixar, and Star Wars. A landmark deal at the intersection of AI and entertainment.
Not a single cent ever changed hands.
When OpenAI decided to shut down Sora, Disney learned about it less than one hour before the public announcement, according to the Wall Street Journal. Disney’s official response was diplomatically ice-cold: “We respect OpenAI’s decision to exit the video generation market and refocus its priorities.”
This isn’t just an embarrassing anecdote. It’s a warning signal for the entire AI partnership ecosystem. When the underlying unit economics don’t work — $1.30 per 10-second clip with no viable monetization model — even the most prestigious partnerships can’t save the product.
Disney was banking on Sora’s consumer reach; OpenAI was banking on Disney’s intellectual property to reignite engagement. Both bets fed off each other — and collapsed together.
Why Productive AI Is Winning the War
While Sora was burning $15 million a day, productive AI tools were quietly racking up wins.
Anthropic captured 40% of the enterprise AI API market, compared to 25% for OpenAI, according to Menlo Ventures’ 2025 enterprise AI report. In the specific category of coding, it’s even more striking: Anthropic holds 54% of the market, versus 21% for OpenAI.
Claude Code has become the product that’s “eating OpenAI’s lunch,” in TechCrunch’s words. And it’s precisely to counter this threat that Altman decided to kill Sora: free up the compute for tools that actually generate enterprise revenue.
The economic model of productive AI is fundamentally different:
- Low marginal cost — a code request costs a fraction of what a video generation costs
- Recurring revenue — monthly subscriptions create a predictable revenue stream
- High switching costs — once integrated into a team’s workflow, the tool becomes indispensable
- Provable ROI — companies can measure the productivity gains
This is what JPMorgan identified when the bank described OpenAI’s moat as “increasingly fragile.” The real moat in AI isn’t model quality — it’s workflow integration and user dependency.
And that’s exactly where creative tools fail: they don’t create lock-in. A designer can switch from Midjourney to DALL-E 3 to Flux in a single day. A developer who’s built their entire workflow around Claude Code won’t switch that easily.
Who’s Next on the Chopping Block?
If the creative-vs-productive thesis holds, then Sora is only the first domino.
Runway is best positioned to survive. Its Gen-4 Turbo model is 2.5 times cheaper than Sora on compute, it targets advertising and film professionals (not consumers), and it charges users a price that covers costs. Runway understood it needed to be a B2B tool, not a viral toy.
Kling 2.5 by Kuaishou is playing the high-volume, low-cost game, primarily for China’s social content market. Its edge: lower infrastructure costs in China and a massive domestic market.
Google Veo 3 benefits from the Google ecosystem: in-house infrastructure (TPUs), distribution via YouTube, and the ability to absorb video losses as long as AI boosts ad revenue.
The most vulnerable? Tools that, like Sora, target consumers without a clear monetization model and with high compute costs. Midjourney, despite estimated annual revenue of $200 million (according to The Information), faces growing price pressure from open-source models like Flux and ByteDance’s Seedance.
The pattern is clear: the creative AIs that survive will be the ones that behave like productive AIs — selling professional tools with measurable ROI, not viral meme generators.
FAQ
Will Sora come back in a different form?
OpenAI reassigned the Sora team to a research project called “world simulation” for robotics — an approach closer to LeCun’s world models than to consumer video. The idea is to repurpose video synthesis technology to train robots in simulated environments — a use case where compute cost is an R&D investment, not an operational money pit. But a comeback as a consumer product seems off the table.
Why didn’t OpenAI just raise Sora’s prices?
The problem wasn’t the price — it was perceived value. With 30-day retention below 10%, users didn’t consider Sora useful enough to pay more. Raising prices would have accelerated user flight without solving the fundamental problem: the absence of a recurring use case.
Is AI video doomed?
No, but it needs to change its model. Runway, Kling, and Veo 3 prove that AI video can be viable — as long as it targets professionals, optimizes compute from day one, and charges a price that reflects actual costs. What’s doomed is the “free worldwide demo” model that Sora attempted.
Key takeaways:
- Creative AI burns cash, productive AI generates it. Sora’s cost-to-revenue ratio ($15M/day vs. $2.1M total) is the textbook case of this rift.
- Measurable ROI makes all the difference. Tools embedded in daily workflows (code, enterprise) justify their costs. Viral novelties don’t.
- Compute is the ultimate bottleneck. Altman killed Sora to free up GPUs against the Anthropic threat. In AI, every GPU must earn more than it costs — or it gets cut.
