On March 24, 2026, OpenAI announced the complete end of its Sora video-generation AI. Since appearing as a research preview in February 2024, its extraordinary visual expression had excited creators around the world as something that would rewrite the future of video production.
Yet Sora would disappear from the market only a little more than two years after its spectacular debut. The consumer web and app versions ended on April 26, 2026, and the developer API will be shut off on September 24 after a phased transition. Video generation built into ChatGPT has also effectively stopped.
Why was such a closely watched major project terminated so quickly? The background combines soaring operating costs, the dead end of a subscription business, an unbridgeable gulf with the entertainment industry, and the barrier of copyright law in multiple countries. At a deeper level, the decision was also a ruthless change of direction: OpenAI strategically abandoned the video-generation market itself and rapidly redirected resources toward its next principal arena.
Withdrawal after Only Two Years: From Excitement to Disillusionment
Sora's path mirrors the process by which expectations for generative AI swell and then rapidly contract when they meet business reality. Released in February 2024, Sora shocked the industry with high-quality footage that seemed to understand physical laws. In December, the first model became available to paid ChatGPT users, putting video-generation AI in the hands of the public.
The turning point was Sora 2, a standalone app released at the end of September 2025. Sam Altman positioned it as a milestone comparable to the first appearance of ChatGPT. The app did attract enormous attention, and active users temporarily reached about one million.
The excitement did not last. Long waits for generation and unstable results drove casual users away, and active users soon fell below 500,000. App downloads dropped to roughly one-third of their November 2025 peak by February 2026.
Safety filters were strengthened in November 2025 to address concerns including deepfakes, which restricted uses likely to spread widely. The free allowance quietly disappeared in January 2026, at about the same time a major partnership with Disney was announced. OpenAI sought to abandon free users and shift its center toward large enterprises.
That attempt also failed, and the service closure was formally announced on March 24. Sora's development team was reassigned to robotics and world-model research. Withdrawal only six months after the standalone app's release suggests that internal cash consumption far exceeded initial assumptions.
The Economics Did Not Work: Video Generation as a “Computing-Resource Quagmire”
The most direct reason Sora abandoned the business was that its economics had fundamentally collapsed. Applying the business model established for text generation directly to vastly heavier video generation encountered the limits of hardware and enormous server costs.
Generating video requires incomparable amounts of computation relative to text. Producing high visual quality can involve tens or hundreds of times more data than a still image, and quality-oriented Sora repeated that processing through additional layers.
Text generation can release computing resources as soon as output ends, but one video request occupies high-performance GPUs for several minutes. Users paid a fixed monthly price while every use increased costs, creating a structural defect in which adding users increased losses.
Reports put Sora's daily operating cost at around $1 million, or approximately ¥150 million, and say that server costs alone reached as much as $15 million per day during periods of concentrated demand.
Despite this expenditure, Sora reportedly earned only about $2.1 million in revenue across its entire life. Some estimates placed the effective generation cost above $1 for a single ten-second video.
The burden also weighed heavily on OpenAI's overall finances. The company was already recording enormous losses, and continued operation of a costly model was reportedly expected to deepen the deficit.
As an initial public offering approached and the company needed to demonstrate financial soundness to investors, separating itself from unprofitable Sora became unavoidable. The greatest force behind the closure was the judgment that scarce, valuable GPUs should go to more profitable coding assistance and enterprise services rather than entertainment.
Why the $1 Billion Disney Partnership Collapsed
The most dramatic illustration of Sora's limits was the collapse of a major Disney partnership valued at $1 billion. Led by former CEO Bob Iger, the multiyear licensing agreement was concluded at the end of 2025. The plan was to integrate more than 200 characters from Disney, Marvel, Pixar, and Star Wars into Sora, allowing officially licensed AI video generation, and there was discussion of bringing generated character videos into Disney's streaming service.
On March 24, 2026, just before announcing Sora's end, Altman directly called Josh D'Amaro, who had become Disney's new CEO only on March 18, and informed him that the project and partnership would be terminated. Notice only 30 minutes before the announcement sent shock waves through the industry. Disney said it respected OpenAI's changed priorities, but the agreement, including investment, effectively returned to a blank slate. Under D'Amaro, Disney reportedly changed direction and began negotiations with other AI companies.
At the root of the collapse was a more fundamental disagreement than scheduling or funding. Professional video production demands exact frame-level control. Camera movement, lighting color, character expressions, and mouth motion must all be managed numerically and precisely, with the same result reproducible as many times as needed.
Sora's mechanism, however, probabilistically inferred the most plausible footage from submitted text. Even if a director gave mathematically precise instructions, the structure could not return the same intended result reliably. Disney's technical team judged it difficult to integrate probabilistic Sora into existing production and saw too much risk of brand damage from opening valuable characters.
The Copyright Barrier: What Japan's CODA Confronted
Alongside business problems, legal and ethical risks inherent in video-generation AI made continued service difficult. Because Sora could easily create video closely resembling existing work and real people, it remained caught between users' creative freedom and safe operation.
After the public app release, social media filled with unauthorized imitations of anime characters and scenes from famous works. Large numbers of videos inappropriately depicting historical figures and nonconsensual deepfakes also drew strong criticism. OpenAI had to strengthen its safety filters substantially, creating a vicious cycle that reduced expressive freedom and weakened the tool's appeal.
A public organization in Japan also expressed formal concern. The Content Overseas Distribution Association (CODA) confirmed that Sora 2 generated many videos closely resembling well-known Japanese anime works and characters and submitted a request to OpenAI dated October 27, 2025. It called for an end to unlicensed training and a sincere response to infringement.
CODA stated that where particular copyrighted works are reproduced or generated in similar form, copying during training itself could constitute copyright infringement. It argued that Japanese law requires advance permission in principle and has no system allowing responsibility to be escaped through objections after the fact.
OpenAI subsequently reportedly informed CODA that Sora 2 would end. The episode became symbolic of the difficulty of globally commercializing a general-purpose generative AI across the complex copyright laws of individual countries.
The Real Objective Was a “World Model”
Each factor above triggered the closure, but remains a surface-level reason. The real reason OpenAI released Sora at this point was a major company-wide shift toward developing “world models” and AI operating in physical space: robotics.
Unlike a language model that learns patterns in text, a world model understands from within the laws and causal relationships of the physical world—gravity, reflections of light, collisions between objects, and the passage of time—and predicts what will happen after an action.
If existing AI is an excellent office worker organizing and outputting information, a world model is an evolution toward a “brain” that understands its environment and acts while anticipating what comes next.
Through Sora's development, OpenAI reached an important fact: to generate realistic video, AI itself must fundamentally understand the passage of time, motion of objects, and causality.
Sora's video generation was therefore both an entertainment tool and an attempt to build a simulator of the physical world within the model. From this perspective, Sora's public release can be seen not merely as an app launch but as an enormous data-collection experiment to teach AI the physics of the real world.
Once gravity and movement had been learned to a certain level, Sora's role was complete. OpenAI abandoned the loss-making consumer app and absorbed the technology and development team supporting it back into the company.
The large volume of GPUs freed by the closure was concentrated on final adjustments to a next-generation model under the internal code name Spud. It was released on April 23, 2026, as GPT-5.5 and is expected to form a foundation for more advanced autonomous AI.
Once world models begin operating in earnest, AI's impact will extend beyond digital space. Expected uses include reading machine vibrations in a factory to predict a failure and order parts in advance, or reorganizing freight loading around weather and traffic constraints.
Flexible movements also come into view, such as predicting and supporting an older person's fall in care settings or grasping a fragile object with exactly the right force. OpenAI concluded that the next enormous market lay not in digital entertainment but in the physical world and AI controlling it, and bet on concentrating resources there.
Where Will Leadership in Video-Generation AI Go? Market Realignment
The withdrawal of absolute pioneer Sora does not mean the end of the video-generation AI market. Rather, it signals that the stage of showcasing technology has ended and serious competition for practical implementation and survival has begun.
Major Chinese platform companies with their own economic ecosystems are rapidly filling the space Sora left. Companies such as ByteDance and Kuaishou contain enormous domestic markets for short video and ecommerce and have the advantage of connecting AI directly at their centers. ByteDance's Seedance earns high praise for cinematic visual quality, while Kuaishou's Kling excels in natural human movement and emotional expression and has spread widely, especially in Asia.
Where Sora prioritized only quality improvements and deferred economic viability, these companies emphasize viable cost structures by making models lighter and using their own chips.
Yet technically advanced Chinese companies face the same or even higher copyright barriers as Sora when expanding globally. ByteDance's Seedance 2.0 symbolized the problem. Though praised for strong performance, it received warnings from Hollywood companies including Disney and Paramount just before global expansion. Strong complaints over unauthorized training forced deployment to stop.
This clearly showed that future market leadership will require more than sustainable computing infrastructure and monetization. It will also demand a clean-data strategy that clears complex international copyright requirements.
What Sora Left Behind
Sora's end in March 2026 cannot be dismissed as the collapse of an AI bubble or one product's defeat. It represents a strategic management decision that ruthlessly calculated how to use enormous computing resources and immediately cut an unprofitable business.
The Sora project showed the entire industry how heavy a burden video generation places on current semiconductors. At the same time, it produced the most important by-product on the path to artificial general intelligence: teaching AI the physical laws of reality.
The app named Sora disappeared in less than two years. Yet the shift in thinking it brought to global AI research and industry, and the enormous physical-simulation data extracted from it, will live on in changed forms as a foundation for the coming “physical artificial general intelligence.”
