OpenAI Sora: The Rise and Fall of the AI Video Generator That Burned $1 Million a Day

When OpenAI previewed Sora on February 15, 2024, the demo reel did something rare for an AI launch: it made people who were tired of AI demos sit up. A woman walked through a neon-lit Tokyo street, her coat and the puddles and the reflected signage all behaving roughly the way light and fabric behave. Woolly mammoths kicked up snow. A flock of paper airplanes moved through a forest. The clips ran up to a minute at high resolution, and nothing else on the market in early 2024 was close. Runway and Pika were producing three-second curiosities. Sora produced something you could mistake, for a few seconds, for footage. OpenAI did not ship it. It posted the reel, named the model, and went quiet, which only sharpened the wanting.

The wait ran most of a year. Sora opened to ChatGPT Plus and Pro subscribers in the US and Canada on December 9, 2024, behind subscription tiers and, at times, behind a closed door: OpenAI paused new signups within hours of launch when demand overwhelmed capacity. The gap between the February reel and the December product was the first crack. The cinematic demos had been the best outputs of a research model with unlimited patience. The shipping product gave ordinary users a generation that was often good, sometimes uncanny, and unpredictable about which it would be. Hands melted. Objects passed through each other. A gymnast might sprout an extra limb mid-routine. Physics held until it didn't.

Creators who had imagined a one-prompt path to usable video found something closer to a slot machine. Getting a clean ten seconds meant generating, discarding, and regenerating, and the cost of that iteration fell on the user's subscription and patience. The model was impressive in aggregate and frustrating in particular. For hobbyists it was a marvel. For anyone trying to produce something specific, on a deadline, to a brief, the inconsistency was the product's defining trait. Outputs carried a visible watermark, clips topped out well under the minute the demos had teased, and the monthly credit allotment drained fast for anyone generating in bulk. Access was gated by price as well as quality, twenty dollars a month for Plus and two hundred for Pro, and even the Pro tier capped how much video you could make. The waitlist had promised a tool. What arrived was an experiment you paid to run.

OpenAI's answer, in late September 2025, was not a quiet fix but a bigger bet. Sora 2 arrived on September 30 with a standalone iOS app built around a social feed, a vertical scroll of AI-generated clips with a cameo feature that let people drop themselves and their friends into generated scenes. An Android version followed two months later. This is the version most people picture when they picture Sora, and it is the version that died. Sora 2 was not a successor that solved the original's problems. It was the product OpenAI shut down.

The numbers underneath the app explain why. Sora's audience peaked near one million users and then slid below half a million as the novelty wore off and the feed filled with the same gags. Reported 30-day retention for paying Pro users sat under eight percent, which is to say more than nine in ten of the people most likely to pay had stopped showing up within a month. Against that, OpenAI was reportedly spending close to a million dollars a day to run the model. The cost of generating video is brutal in a way that text generation is not: every second of output is a small render farm's worth of compute, and some reports put Sora's inference bill far higher than a million a day. The revenue to cover it never arrived.

While Sora burned, the field it had defined caught up and undercut it. Runway shipped successive Gen models with the frame-level control that working editors actually wanted. Pika kept iterating cheaply. Google's Veo matched Sora on quality from inside a company with its own data centers. And a wave of Chinese models, Kuaishou's Kling foremost among them, reached comparable output at a fraction of the price. Sora's eighteen-month head start in mindshare did not convert into a durable lead in product. By 2026 the question was no longer whether you could generate a convincing clip. It was who could do it cheapest, with the most control, at a price someone would pay. Runway and Luma raised on the same thesis and kept shipping, while the open release of capable Chinese models meant a hobbyist could match much of Sora's output for nothing. The novelty that had carried Sora's launch was, by then, a commodity.

The strangest chapter came near the end. In December 2025, Disney announced a roughly one-billion-dollar arrangement with OpenAI: an equity investment paired with a licensing deal that would have let Sora users generate short videos featuring more than two hundred characters from Disney, Marvel, Pixar, and Star Wars. It was the first time Disney had licensed its characters to an AI platform, and on paper it gave Sora the one thing it lacked, a reason to keep coming back. Three months later it was gone. No money had changed hands and no formal agreement had been signed when OpenAI decided to wind Sora down, and the deal's logic collapsed with the product it depended on. Disney walked, with a line about the future being human.

OpenAI announced the end on March 24, 2026, in a post on X. The public framing was a refocus: the company was concentrating on its core products and reallocating compute toward the work that paid, chiefly its enterprise and coding business. The Sora app went dark on April 26, 2026, and the API is set to follow on September 24. The reporting around the decision was blunter. Sora was consuming GPUs OpenAI needed elsewhere, its retention and revenue did not justify the spend, and the company's priorities had moved toward enterprise tools and, by some accounts, robotics. Bill Peebles, who had led Sora since its inception, left OpenAI within weeks of the announcement. His co-lead Tim Brooks had already departed for Google in late 2024. Users were given a window to download their generations before the lights went out, the standard sunset courtesy, and then the most talked-about AI product of early 2024 was a redirect.

OpenAI was careful to say it was not leaving video generation, only retiring the Sora brand and its apps. A replacement model was reportedly in development, and video features were expected to surface inside ChatGPT, though even that was muddled: some reports indicated ChatGPT would not carry video either, at least not soon. Whatever arrives next will be a different product under a different name. Sora, the standalone app and the brand that briefly owned the category, is finished.

Which leaves the question the whole episode poses: is text-to-video simply hard to monetize, or did OpenAI just run this one badly? The structural case is strong. Video inference is expensive in a way that scales with quality and length, the consumers who will pay are few, and the professionals who would pay want control and consistency that diffusion models still deliver unevenly. A feed of AI clips is cheap to scroll and expensive to serve, which is the wrong shape for a business. The execution case is also fair. OpenAI chased a social app rather than the creative tooling editors asked for, and it did so as a company increasingly pointed at enterprise revenue, where a consumer video toy was always going to lose the fight for compute. Both readings can be true. The survivors, Runway and Kling and the rest, are not winning because they cracked the economics. They are winning because they cannot afford to stop trying, and OpenAI could.

Sora's failure is less a verdict on AI video than on the distance between a demo and a business. The February 2024 reel was real, and it was ahead of everything else. What OpenAI never built was a reason for the people who paid to keep paying, or a cost structure that survived their leaving. The product that launched the category was killed by the category's basic economics: a model too expensive to run for an audience too thin to fund it, inside a company with more profitable places to point its chips. The lesson for the next launch is the oldest one in this archive. The demo is not the product, the product is not the business, and the company that ships first is rarely the one that has to make it pay.