In late September 2025, OpenAI launched an app that shot to number one on the US App Store faster than ChatGPT ever did.

Six months later, OpenAI killed it.

Not because it flopped at launch. It was one of the most talked-about apps of the year. People made videos of themselves as action heroes, dropped their friends into film scenes, and shared them like TikToks.

The problem was simpler and scarier than a flop.

Sora cost OpenAI around $1 million a day to run. Over its entire life, it made about $2.1 million in total. That is roughly two days of revenue to cover six months of bills.

This is the story of how a genuine hit became too expensive to keep alive. And why the smartest product decision OpenAI made all year was to switch off its own viral app.

"When I think of revenue, I think of Attio." - Shreman Shrestha, Head of Business at Granola

Here's what that adds up to:

Start Your Trial

Sora did not start as an app.

It began as a text-to-video model. OpenAI first showed Sora off in February 2024, then quietly launched it in December 2024 for paying ChatGPT users in the US and Canada. You typed a sentence, and it made a short video. Impressive, but niche.

The real moment came on 30 September 2025.

That day, OpenAI released Sora 2, a much better model, and wrapped it in a brand-new social app. TechCrunch called it OpenAI's own TikTok competitor.

The app had a scrolling video feed. Everything in it was AI-generated.

The killer feature was called cameos. You recorded your face and voice once, and then you could drop yourself into any scene. Your friends could use your likeness too, with your permission.

It spread fast.

The app was free and invite-only at first. That mix of "free" and "you need an invite" created the same fear-of-missing-out that made Clubhouse explode years earlier. Everyone wanted in.

Within weeks, Sora passed one million downloads faster than ChatGPT had. Downloads peaked at over 3.3 million in November 2025.

For a moment, it looked like OpenAI had cracked social media on its first try.

Three things made Sora take off.

First, the technology felt like magic. Sora 2 followed the laws of physics better than anything before it, and it added sound. A person could speak, and the lips matched. That "how is this real?" feeling is what made people share.

Second, cameos made it personal. Most AI video tools make clips of strangers. Sora made clips of you. Personal content gets shared, and shared content brings new users. That is a growth loop most apps would kill for.

Third, it had OpenAI's name on it. ChatGPT already had hundreds of millions of users. Sora borrowed that trust and attention on day one.

Here is the trap hiding inside all of this.

Every one of those strengths made the app more expensive to run. More sharing meant more videos. More videos meant more of OpenAI's most costly resource: computer power.

When your growth loop and your cost curve are the same line, growth is not a win. It is a bill.

In December 2025, Sora looked unstoppable.

On 11 December, Disney and OpenAI announced a landmark deal. Disney would invest $1 billion in OpenAI and license more than 200 characters, from Marvel and Pixar to Star Wars, for people to use in Sora.

This was huge. It was the first big Hollywood licensing deal for AI video. CNBC and others covered it as a turning point for the whole industry.

From the outside, it read as validation. The most protective brand in entertainment had just bet on Sora.

But under the surface, the numbers were already falling apart.

Making AI video is one of the most expensive things you can do with computers.

Text is cheap. A chat reply uses a sliver of power. Video is another world. Estimates suggest a single ten-second clip could tie up several high-end chips for the better part of an hour.

Now multiply that by millions of users making millions of clips, most of them just for fun.

That is how you get to a reported cost of around $1 million per day, according to Wall Street Journal reporting summarised here.

The app was free. OpenAI's only real plan was to charge people for extra videos when demand was high. That brought in almost nothing next to the cost of running the thing.

A product where each use costs you money and earns you nothing does not get healthier as it grows. It gets sicker. Scale is only your friend when the maths per user works. Here it never did.

Then the users started leaving.

This is the pattern that quietly kills "viral" apps. Getting people to download something is easy when it is new and everyone is posting about it. Getting them to come back next week is the hard part.

Sora was built on novelty. Once you had made a few clips of yourself as a superhero, the magic faded. There was no deep reason to return every day.

The numbers show it plainly.

After the November peak, downloads fell 32% in December, then 45% in January. By February 2026 they were down to about 1.1 million a month, a 66% drop in three months.

In the App Store, Sora slid from number one all the way to number 101.

Active users told the same story. They peaked near one million, then fell to under 500,000 before the shutdown.

Downloads are a vanity metric. Retention is the truth. A million people trying your app once is a party. A hundred thousand coming back every day is a business. Sora had the party, not the business.

Sora also had a content problem, and it cut both ways.

Early on, people used it to make disturbing deepfakes. The app generated videos of real dead figures like Martin Luther King Jr and Robin Williams, which upset families and drew heavy criticism.

So OpenAI tightened the rules.

But tighter rules frustrated the people who came to play. Copyright limits meant users could no longer make videos of their favourite characters freely. The very thing that made Sora fun was now fenced off.

That is the hard squeeze of a user-generated content product. Too loose, and you become a deepfake machine that scares off partners and regulators. Too tight, and you bore the users who made you popular. Sora never found the middle.

By early 2026, Sora was no longer the only option. Google and Meta were pushing their own AI video tools hard.

More importantly, OpenAI's own priorities had shifted. The company was pouring compute into ChatGPT and its enterprise products, the things that actually made money. Video was eating chips that were needed elsewhere.

When a resource is scarce, every product competes for it. Sora was losing that fight against OpenAI's core business, and it deserved to.

On 24 March 2026, OpenAI announced it was shutting Sora down.

The app and website would go dark on 26 April 2026. The developer tool, or API, would follow on 24 September 2026.

The Disney deal collapsed with it. Remember that $1 billion partnership from December? Reports say Disney found out about the shutdown less than an hour before it went public. No money had actually changed hands.

It is easy to call this a failure. In one sense it was. A viral app with a Disney deal died in six months.

But look closer, and it is also discipline.

OpenAI had a product that was loved, famous, and bleeding money with no path to profit. Plenty of companies would have kept it running for the headlines, hoping the maths fixed itself. It rarely does.

Knowing when to kill a product is a product skill too. Every day Sora ran, it burned cash and compute that could go to work that paid off. The brave call was not launching the shiny app. It was switching it off.

The Sora app is gone. But the technology is not.

The Sora 2 model still lives on inside ChatGPT, behind a paywall. Paying users can still make videos. What died was the free, social, all-you-can-eat version that anyone could pile into.

That tells you what OpenAI really learned.

The model (that generated videos) was never the problem. The business model was. A powerful tool sold to people who pay for it can work. The same tool given away to millions for fun cannot, at least not yet, when every video costs real money to make.

Sora was not a bad product. It was a great demo wearing the costume of a business.

The lasting question for the rest of us is the one every "viral" launch should ask on day one: if this takes off, does each new user make us stronger or poorer? For Sora, growth was the disease, not the cure.

Would you have pulled the plug as fast as OpenAI did, or ridden the hype a little longer? Reply and tell me where you land.

If you found this useful, forward it to a PM who is chasing downloads instead of retention. That is the whole game.

Until next time,
—Sid