In the end, Sora survived barely sixteen months in public hands. OpenAI pulled the plug on the text-to-video app April 26, 2026. The company plans to kill the API on September 24, 2026. That date will mark the formal death of the whole project.
What killed it? A watermark that lasted one week.
The first generation of Sora hit ChatGPT Plus and Pro subscribers in the U.S. and Canada on December 15, 2024. It could take a text prompt and spit out a short video clip. It could also extend existing short videos. OpenAI had previewed sample output back in February 2024, so nobody was blindsided by what the model could do.
But from the start, the company had a problem. By default, Sora trained on copyrighted material. It used that material in the videos it generated. The burden fell on copyright holders to opt out. That arrangement angered artists, studios, and stock footage companies almost immediately.
OpenAI’s answer was a visible, moving digital watermark embedded in every Sora video. The idea was simple: anyone watching could tell the clip was AI-generated. That might have worked, for a while.
It didn’t even last a week.
When Sora 2 launched at the end of September 2025 — a major update that added social media features and opened access to a broader set of select users in the U.S. and Canada — third-party programs appeared almost instantly. Those programs stripped the watermark out. The safeguard became useless.
That was the turning point. A video-generation tool with no reliable way to mark its own output is a liability. It can be used to create convincing fakes with no fingerprint. The technology itself had moved fast. The protections had not.
Sora 2 did integrate social features. It let users interact with the model in new ways. OpenAI called it a significant update. It showed the company’s commitment to improving the technology. But the watermark problem never got fixed. No second-generation safeguard replaced the broken one. The app kept running for another seven months after Sora 2 landed, then shut down.
The API will follow this September. After that, no part of the Sora brand remains.
What stands out about the timeline is how compressed it is. Preview in February 2024. Public release in December 2024. Major update in September 2025. Shutdown in April 2026. The whole arc from debut to burial took roughly two years and two months. For a product that represented a milestone in AI development, the run was brutally short.
OpenAI never commented publicly on why Sora died. But the facts of the case tell their own story. A watermark that failed in a week. Copyright holders forced to opt out of a system that used their work by default. A social media update that expanded the user base without solving the core vulnerability. Third-party tools that erased the only identifier the company had built.
Sora was not a niche experiment. It was a major product release from the most visible AI company in the world. It had paying subscribers on two tiers. It had a second generation with new features. And it still could not survive a single week with a working content-authentication system.
The API sunset in September closes the book. There will be no Sora 3. No watermark fix. No opt-in copyright framework retrofitted onto the model. The project is done.
Sixteen months of public service. One week of working protection. That is the math behind Sora’s end.
