In February 2024, OpenAI unveiled Sora, a text-to-video model that generated short, high-definition clips from simple prompts. The demos—an SUV winding down a mountain road, people walking through snowy Tokyo, whimsical creatures interacting with their environment—looked remarkably coherent and cinematic. Within hours, the internet declared that the future of video had arrived. Less than two and a half years later, the company that created it shut the product down.
The Sora web and app experiences ended on April 26, 2026. The API is scheduled to follow on September 24, 2026. What began as one of the most hyped AI demonstrations of the decade became a case study in the brutal economics, technical limits, and strategic realities of generative video.
The Dramatic Rise
OpenAI first publicly demonstrated Sora on February 15, 2024. CEO Sam Altman shared generated clips on X in response to user prompts, and the company published a technical overview. Sora was described as a diffusion transformer model that treated video as sequences of “spacetime patches.” It could produce videos up to roughly one minute long while attempting to maintain visual consistency, camera movement, and basic physical coherence—capabilities that earlier systems such as Runway’s Gen-2 struggled with.
Access remained tightly restricted for most of 2024. Red-teamers tested for safety and misuse; a small group of creative professionals received early access. Public availability arrived on December 9, 2024, for ChatGPT Plus and Pro subscribers in the United States and Canada. Users could generate short clips, extend existing videos, and iterate on prompts. The quality was uneven—complex physics, precise object interactions, and long-term temporal consistency remained weak—but the leap over previous public models was undeniable.
The real product expansion came with Sora 2 on September 30, 2025. OpenAI launched a dedicated iOS app (Android followed later) that combined improved generation quality with social features. The experience resembled a short-form video platform: users created clips, shared them, and browsed others’ work. Visible watermarks were added to reduce the risk of undetectable deepfakes. At its peak, worldwide users reached around one million before declining to fewer than 500,000.
In December 2025, OpenAI and Disney announced a landmark three-year partnership. Disney committed approximately $1 billion and licensed more than 200 characters from Disney Animation, Pixar, Marvel, and Star Wars for use inside Sora. The deal was widely viewed as a potential bridge between Hollywood and frontier AI companies. It lasted only a few months.
Technical Achievements and Persistent Limits
Sora’s core innovation lay in scaling diffusion models to video while preserving more temporal consistency than earlier systems. It could generate different camera angles of the same scene without explicit prompting and produced surprisingly coherent 3D-like motion from pure text. Training reportedly used a combination of publicly available video and licensed material, with recaptioning techniques (similar to those used in DALL·E 3) to improve prompt adherence.
Yet the model never fully solved fundamental problems. Complex physical interactions often failed. Causality and object permanence broke down over longer sequences. Distinguishing left from right remained unreliable. Safety filters blocked sexual, violent, and certain celebrity content, but enforcement was imperfect. Deepfake videos of deceased celebrities and public figures still appeared, prompting legal threats from estates and advocacy groups.
Cultural and Industry Impact
Sora’s early demos accelerated investment and research across the industry. Competitors including Runway, Kling, Luma, and Chinese models such as Seedance intensified development. Filmmakers and studios experimented with concept work, previz, and short-form content. At the same time, anxiety ran high. Filmmaker Tyler Perry publicly paused an $800 million studio expansion, citing the potential disruption AI video tools could cause to traditional production jobs. Industry analyses warned that generative AI could affect tens or hundreds of thousands of entertainment-sector roles in the coming years.
Copyright became a flashpoint. Sora’s default approach allowed generation of copyrighted styles and characters unless rights holders actively opted out. Japan’s Content Overseas Distribution Association and major Hollywood guilds (SAG-AFTRA, Writers Guild of America, Animation Guild) raised strong objections. The Disney deal itself drew criticism from creative unions who described it as legitimizing the “theft” of their work. OpenAI later promised more granular controls for rights holders and explored revenue-sharing models, but these efforts arrived late.
On the positive side, the technology lowered barriers for independent creators, educators, and marketers. Short educational explainers, product visualizations, and experimental narrative pieces proliferated. Academic work noted that generative video tools expanded participation in creative production for people outside traditional industry pipelines.
The Economic Reality That Ended It
By early 2026 the numbers no longer worked. According to reporting that cited internal figures and Wall Street Journal investigation, Sora’s worldwide active users peaked near one million and then fell below 500,000. Operating the service cost an estimated $1 million per day because of the extreme computational demands of high-quality video generation. Lifetime in-app purchase revenue was reported in the low single-digit millions—figures around $1.4 million to $2.1 million depending on the source—orders of magnitude below the cost of running the platform.
Video generation is simply expensive. Each short clip required substantial GPU time. Sustained consumer usage at scale, especially on a free or lightly monetized social product, created a structural loss. Meanwhile, OpenAI faced intensifying competition in other domains. Coding agents and enterprise tools from rivals such as Anthropic were gaining traction among the developers and companies that form OpenAI’s most important commercial base.
On March 24, 2026, OpenAI announced it was discontinuing Sora. An official post stated: “We’re saying goodbye to Sora. To everyone who created with Sora, shared it, and built community around it: thank you.” The company told the BBC it was ending the product so it could focus on other developments, including robotics “that will help people solve real-world, physical tasks” and agentic systems capable of more autonomous work. The same underlying video understanding research would be redirected toward training robots that operate in the physical world. Image generation inside ChatGPT was unaffected.
The Disney partnership was dissolved. No money had reportedly changed hands. Disney stated that it respected the decision and would continue exploring responsible AI applications with other partners.
Strategic Pivot and Broader Lessons
The shutdown fits a larger pattern. OpenAI has repeatedly pruned products and features that do not clearly advance its core trajectory toward more capable general systems and commercial enterprise use. High compute costs, modest consumer retention, ongoing safety and copyright liabilities, and the opportunity cost of diverting GPUs from higher-priority research all pointed in the same direction.
Sora’s brief life demonstrated both the power and the fragility of frontier generative models. The technology proved that coherent, prompt-driven video generation was possible far sooner than many expected. It also showed that quality alone does not guarantee a sustainable product when inference costs are extreme, monetization is weak, and legal and social risks remain high.
Competitors continue to improve their own systems. Some focus on professional tools with tighter controls and clearer licensing; others pursue consumer applications with different economics. The research insights from Sora—particularly the ability to simulate aspects of the physical world—will likely live on inside OpenAI’s robotics and world-model efforts.
For creators who used the tool, the discontinuation was abrupt. OpenAI provided a data export window so users could download their generations before permanent deletion. The API remains available until late September 2026 for those still integrating it into workflows.
Conclusion
Sora arrived with the force of a genuine technological leap and left as a cautionary tale about the gap between demonstration and durable product. It forced Hollywood, regulators, and the broader public to confront questions about authorship, consent, deepfakes, and the economics of generative media more urgently than any previous AI system.
Its discontinuation does not mean AI video generation has failed. It means that one particular product, built and operated under a specific set of constraints and ambitions, proved unsustainable. The underlying capabilities continue to advance elsewhere. The next generation of video models will be judged not only by how impressive their demos look, but by whether anyone can afford to run them, license them cleanly, and keep them online.
In that sense, Sora’s short, intense life may have been less a failure than an expensive, highly public experiment that clarified the real constraints of the field. The sky it was named after remains open. The path to filling it with useful, reliable, and economically viable generated video is still being mapped.