What Happened to Sora? The Rise and Fall of OpenAI's Video AI

In February 2024, OpenAI showed the world a video of a woman walking through a neon-lit Tokyo street. It was generated entirely by AI. The internet collectively lost its mind. Hollywood panicked. Creators panicked. The phrase "Sora changes everything" appeared in roughly ten thousand LinkedIn posts within 48 hours.

Twenty-five months later, OpenAI shut Sora down. The app closes on April 26, 2026. The API follows in September. The product that was supposed to change everything changed almost nothing — except as a cautionary tale about hype, unit economics, and what happens when you bring a research demo to a product fight.

Here is what actually happened. And why every business that relies on AI tools should be paying attention.

The Timeline — From Demo to Dead

February 2024: OpenAI previews Sora with a series of jaw-dropping demo videos — an SUV on a mountain road, a Tokyo street scene, historical footage of the California gold rush. No public access. Maximum hype. The demos were cherry-picked, hand-curated, and represented the absolute best the model could produce. Nobody knew that at the time.

November 2024: A group of beta testers leak a Sora API key on Hugging Face, along with a manifesto accusing OpenAI of "art washing." OpenAI revokes all access within three hours. The leak reveals that the actual user experience was considerably less polished than the demos suggested.

December 9, 2024: OpenAI launches Sora publicly as part of their "ship-mas" event — a 12-day product unveiling series. Sora Turbo becomes available to ChatGPT Plus and Pro subscribers. Users can generate videos up to 20 seconds at 1080p. Downloads peak at 3.33 million in the first month. The hype machine is running at full speed.

September 2025: Sora 2 launches with an iOS app, followed by Android two months later. It is supposed to be the version that closes the gap with competitors. It does not.

December 2025: Disney commits to a $1 billion investment and a three-year character licensing arrangement with OpenAI — covering more than 200 masked, animated, and creature characters from Disney, Marvel, Pixar, and Star Wars. Sora is positioned as a core part of the partnership. No money changes hands.

February 2026: Downloads have fallen 66% from peak, down to 1.13 million. The product is declining before anyone pulls the plug. Users are discovering that generation times of 3-8 minutes per clip make iterative workflows impossible. Runway and Kling are producing comparable quality in a fraction of the time.

March 24-27, 2026: OpenAI announces Sora's shutdown. The Sora app closes April 26, 2026. The API closes September 24, 2026. Disney learns about the shutdown less than an hour before the public announcement. The $1 billion deal is dead.

The Numbers That Killed Sora

The financials were catastrophic. This was not a product that failed to find product-market fit. It was a product that found an audience and then haemorrhaged money serving them.

Estimated operating cost: $15 million per day. Running video generation at scale requires enormous GPU compute. Each minute of Sora-generated video consumed approximately 10 to 15 times the compute of a standard ChatGPT conversation. OpenAI explored custom ASIC solutions with Broadcom but could not get costs below $1.40 per generated second — still 16 times higher than Kling's API pricing of $0.084 per second. The annualised burn rate exceeded $5.4 billion for GPU rental, power, and inference alone.

Total lifetime revenue: $2.1 million. That is not a typo. The entire revenue Sora generated across its lifetime — from consumer app purchases and subscriptions — was roughly what it cost to run the service for three and a half hours.

Peak users: approximately one million, collapsing to under 500,000. Downloads peaked at 3.33 million in the first month after launch, then fell 66% by February 2026, dropping to 1.13 million. The product was not just expensive to run — it was losing the users it was bleeding money to serve.

The 3-8 minute generation times were the killer. Creators who needed to iterate quickly simply could not use Sora as a production tool. They tried it, found it slow, and moved to Runway or Kling. By the time Sora shipped, Pika was generating clips in 15 seconds and Runway's Turbo mode was delivering in under a minute. Sora was not competing on the same timescale.

What Actually Went Wrong

The post-mortems will be written for years. But three things stand out.

1. The demo-to-product gap was enormous. Sora's February 2024 demos were extraordinary. The actual product, when it shipped ten months later, was good but not extraordinary. The demos had been cherry-picked. Real-world usage revealed inconsistent physics, unreliable motion, and artefacts that the carefully curated preview videos had conveniently avoided. Every AI product has a gap between its best possible output and its average output. Sora's gap was wider than most, and the hype had set expectations at the best-possible end.

2. Speed was a fatal flaw, not a minor inconvenience. Three to eight minutes per 10-second clip sounds tolerable in isolation. In practice, it destroyed the iterative workflow that makes AI video useful. Creative work requires rapid iteration — generate, evaluate, adjust, regenerate. At 3-8 minutes per cycle, a session that would take 30 minutes on Runway took an entire afternoon on Sora. By the time Sora shipped, Pika was generating clips in 15 seconds. Sora was not competing on the same timescale.

3. OpenAI tried to be a consumer app company and discovered it was not one. OpenAI's strength is models and APIs. Its enterprise customers — developers, businesses, platform integrators — pay reliably and at scale. Sora was a consumer product aimed at creators, a market that is notoriously difficult to monetise. The company that makes GPT-4 and powers thousands of enterprise applications decided to also be a video editing startup. It turns out those are different businesses requiring different competencies, different go-to-market strategies, and different patience from leadership.

The Disney Disaster

The Disney deal deserves its own section because it illustrates a broader risk that every business working with AI vendors should understand.

Disney committed to a $1 billion investment in OpenAI in December 2025, with a three-year character licensing agreement covering over 200 characters from Disney, Marvel, Pixar, and Star Wars. Sora would have generated user-prompted videos featuring these characters. Three months later, the product at the centre of that partnership was dead. No money ever changed hands.

Disney learned about the shutdown less than an hour before the public announcement. Bob Iger reportedly found out via a phone call from Sam Altman minutes before OpenAI posted to X. Whether this was strategic or chaotic depends on who you ask, but either interpretation is damaging. A billion-dollar partner finding out about a product discontinuation alongside the general public is not how stable vendor relationships work. Reports indicate the relationship between the two companies has been significantly strained.

The lesson is not "do not invest in AI." The lesson is that AI products — particularly at the frontier — are more volatile than traditional software. Features ship and get deprecated. Products launch and shut down. The company that made the best language model in the world decided that its video product was not worth the compute it was consuming, and killed it in weeks. If Disney can get blindsided, so can your business.

What OpenAI Is Doing Instead

OpenAI has not abandoned video generation entirely. Reports indicate a successor project codenamed "Spud," focused on building world models for enterprise and robotics applications rather than consumer video creation. Spud is expected around July 2026 and will be API-only — no standalone app, no consumer-facing product.

The strategic pivot is clear: Sam Altman made the decision to kill Sora and redirect compute resources to more profitable enterprise AI initiatives — particularly coding tools and the upcoming model codenamed Spud — in preparation for OpenAI's IPO. While OpenAI was burning cash on consumer video, Anthropic was quietly winning over the software engineers and enterprises that drive reliable revenue. The consumer video generation market was expensive, competitive, and unprofitable. Enterprise AI — where customers pay per API call and the unit economics actually work — is where OpenAI sees its future.

What to Use Instead

If you were using Sora, or considering it, here is where to redirect.

For professional, cinematic-quality video: Runway Gen-4.5 is the current market leader. Best-in-class fidelity, excellent character consistency, advanced workflow tools. Starting at $12/month for light use, $28/month for real production work.

For high-volume video at competitive cost: Kling 3.0 offers near-Runway quality at roughly 40% of the price. Excellent motion control, video up to 3 minutes, strong character consistency. Starting at $6.99/month.

For fast social media content: Pika 2.5 generates clips in 15-30 seconds — the fastest in the market. Lower quality ceiling than Runway or Kling, but unmatched for high-frequency social publishing. Starting at $8/month for Standard.

For AI avatar and talking-head content: Synthesia or HeyGen, depending on whether your primary use case is training (Synthesia) or sales (HeyGen).

The Lessons for Every Business Using AI Tools

Sora's collapse is not just a story about one product. It is a template for a risk that every business integrating AI tools will face repeatedly over the next five years.

1. Do not build critical workflows on single-vendor AI tools without an exit plan. If your content pipeline depends entirely on one AI platform, you are one product decision away from rebuilding from scratch. Maintain familiarity with at least two alternatives. The switching cost is a few hours of learning. The cost of being stranded is weeks of disruption.

2. Watch unit economics, not demos. A product that is stunning but unsustainable will not survive. When evaluating AI tools, ask how the company makes money on your usage. If the answer is unclear — or if the pricing seems too good to be true — the product may not be long for this world.

3. Distinguish between research previews and shipping products. Sora's February 2024 demo was research. The December 2024 launch was a product. They were not the same thing, and the gap between them was where the disappointment lived. When an AI company shows you a demo, assume you are seeing the best possible output, not the average output.

4. Speed matters more than peak quality for most business use cases. Sora could produce beautiful video. It just could not produce it fast enough for anyone to use it as a real tool. Most business content needs to be good and fast, not perfect and slow. Optimise for workflow speed, not benchmark scores.

The Bottom Line

Sora will be remembered as the product that proved AI video generation was possible and then proved that being possible is not enough. You also have to be fast, affordable, and sustainable.

The companies that survived the AI video shakeout — Runway, Kling, Pika — understood that a production tool needs to fit into a production workflow. They optimised for speed, cost, and reliability alongside quality. Sora optimised for spectacle, and spectacle does not pay the compute bill.

For businesses, the lesson is straightforward: the AI tools you depend on today may not exist tomorrow. Plan accordingly. Diversify. And never confuse a demo with a product.

Digital by Default helps businesses build AI-integrated workflows that are resilient, multi-vendor, and designed for the real world — not the demo reel. If you are rethinking your AI video stack after the Sora shutdown, [get in touch](/contact).