Sora was OpenAI’s AI video-generation model and, later, a social app built around it — type a prompt, or record a short “cameo” of your own face and voice, and get back a realistic video clip. It launched as a research preview in February 2024, grew into a viral, TikTok-style app with the Sora 2 release in September 2025, and was wound down within about a year: OpenAI closed the consumer app on April 26, 2026, and retired the developer API on September 24, 2026. The shutdown followed months of controversy over deepfakes of real people, alongside a strategic pivot at OpenAI toward coding tools and enterprise products.
What Sora was and how it worked
Sora was a diffusion model built on a “diffusion transformer” architecture: it started from visual noise and refined it, step by step, into a coherent video that matched a text prompt, drawing on the same transformer building blocks used in large language models. OpenAI first showed Sora in February 2024, demonstrating minute-long clips of city streets and imagined scenes; ChatGPT Plus and Pro subscribers in the US and Canada got hands-on access in December 2024.
The bigger shift came with Sora 2 in September 2025. It added synchronized dialogue and sound effects, noticeably more realistic physics, and — most consequentially — a standalone iOS and Android app with a TikTok-like feed of AI-generated clips. Its signature feature, “cameos,” let a user record a short one-time video of their face and voice so their likeness could then be dropped into any AI-generated scene, with the same option extendable to friends. A visible moving watermark was meant to mark videos as AI-made, though tools to strip it appeared within days of launch.
Why OpenAI pulled the plug
Two pressures converged. The first was safety and reputational: Sora 2’s realism and the cameo feature made it easy to generate deepfakes of public figures without consent, and OpenAI had to repeatedly intervene after depictions of figures such as Michael Jackson and Martin Luther King Jr. drew complaints from estates and actors’ unions. Advocacy groups and researchers warned more broadly about non-consensual imagery generated from little more than a text prompt.
The second was strategic. OpenAI said it wanted to redirect compute toward coding tools and enterprise customers, and to fold video generation into a broader consolidated product rather than run it as a standalone consumer app. The wind-down also unwound a reported $1.43 billion content-licensing deal with Disney that had been announced only months earlier, with a person close to the deal describing the abrupt cancellation to reporters as a “rug-pull.” OpenAI’s own statement on the shutdown was brief: “What you made with Sora mattered, and we know this news is disappointing.”
What happens to the technology now
Sora is gone as a consumer product, but OpenAI has said the underlying research continues internally, redirected toward “world models” — AI systems meant to simulate physical reality rather than just generate video clips — as part of the company’s stated ambition of eventually applying AI to physical, real-world tasks. Whether that research resurfaces as a public product is, for now, unannounced.
Sora’s rise and fall is also a preview of a broader pattern in AI video generation: tools built by companies like OpenAI can improve rapidly in realism while running well ahead of the guardrails needed to stop them being misused for deepfakes — a gap regulators and platforms are still working out how to close.