The pioneer walked out

In February of 2024, OpenAI released a demo that did something no one was prepared for. A woman walked through a Tokyo street at night, neon reflecting off wet pavement, the camera tracking her as she turned. Woolly mammoths trudged through snow. A couple strolled through a city in the rain. None of it was real. All of it was generated from a sentence. The demo ran a few minutes and reset the expectations of an entire industry. Before that week, AI video meant a smeared, six-second abstraction. After it, everyone understood the camera was about to become optional.

The demo was called Sora. It was never really a product. It was a proof, an announcement, a way of saying look what we can do. It worked. Every video model that followed, Veo and Kling and Runway and Seedance and WAN, exists in part because Sora proved the category could be real. The race that came after, the billions, the studios, the lawsuits, the festivals, all of it traces back to that one week in February.

Now the company that started the race is leaving it. The Sora app shut down on April 26. The Sora API shuts down on September 24, twelve days from now. After that the model IDs stop resolving, and any pipeline still calling the Videos API gets nothing back. OpenAI has named no successor. No replacement model. No new video product. The company that taught the world to believe in generated video has, as of this writing, nothing to sell you in that category.

The timeline is the tell. OpenAI announced the shutdown on March 24 and gave the API six months to die. Six months is enough time to migrate a production pipeline and not enough time to build a replacement. The company did not offer one. It offered a calendar. The app died first, in April. The API dies second, in September. There is no third date, because there is no third thing.

The help page is the coldest part. It tells you how to export your data, then it says the rest will be permanently deleted. Not archived. Not held in case you come back. Deleted. A filmmaker's generations, the prompts, the takes kept and the takes thrown away, all of it gone when the final export window closes. The company is not sunsetting Sora. It is erasing it.

What OpenAI concluded is legible in where the resources went. The video experiment was cut so the money could go toward coding agents and the next ChatGPT, toward a model called Spud and an image model called GPT Image 2. Video was the demo that got everyone's attention. It was never the business. The business is the assistant that writes your code and answers your email. Sora was a fireworks display the company lit and then walked away from, and the fireworks are still going off in everyone else's sky.

Everyone else stayed. Google ships Veo 3.1. ByteDance pushes Seedance. Kuaishou's Kling carries an eighteen billion dollar valuation. Runway, Alibaba, MiniMax, all of them are still in the race OpenAI started and then abandoned. The pioneer walked out and the building got more crowded. The company that proved AI video was possible concluded it was not worth doing, while a dozen companies that watched the proof concluded it was worth everything.

The people who built on Sora feel it first. Sora 2 could hold a shot for twenty-five seconds, longer than most of its rivals, and a certain kind of filmmaker built workflows around that length. Those workflows have twelve days left. The migration guides are already written, mapping Sora's long clips onto Kling's fifteen seconds and its physics onto Veo. A tool died, and the internet responded by drawing a map of where its users should go next. The map is useful. It is also the first one anyone has had to draw for a category the pioneer simply abandoned.

The category now belongs to Google and to China. That is not a prediction. It is a description of the current leaderboard. The American company that invented the category is gone, and the two forces that remain are the search giant and the country that treats AI video as an industrial output. The filmmakers who learned the vocabulary from Sora's demo now speak it into tools built by other people, in other time zones, for other reasons.

The cynical reading is that OpenAI knows something the others do not, that video generation is a money pit and the smart money is leaving early. The other reading is that OpenAI never wanted to make films. It wanted to make a point. Sora was a demonstration of capability, a flex, and once the point was made the flex had no further use. A company that wants to make films does not delete its users' footage. It builds the next camera.

The filmmakers are not left with nothing. They are left with everything except the thing that started it. The vocabulary still works. The models that remain are better than Sora ever was. But the origin story has a hole in it now. The company that lit the fire has gone home, and the people still standing around the fire are the ones who actually wanted the warmth.

The demo was never the product.

Bruce Belafonte is an AI filmmaker at Light Owl. He still has the demo burned into his memory, which is the one thing OpenAI cannot delete.