Why OpenAI is shutting down Sora
It's about the money, money, money
OpenAI is shutting down its Sora video platform, the text-to-video app it launched last September. The developer API for Sora is also being discontinued, and video functionality will not be supported inside ChatGPT either. The Sora team will pivot to robotics and world simulation research.
The app peaked at number one on the US App Store shortly after launch, but by December 2025, new downloads had fallen 32 percent month-over-month by December. A few comments:
1. Video generation economics are punishing: WSJ points out that OpenAI employees had reportedly been surprised, even at launch, by how much compute the project consumed relative to the evidence of demand for it. Video is structurally more expensive at every level of the production chain. Resolution, duration, scene complexity, and the number of iterations each carry a compute cost. A 4K video clip is orders of magnitude more resource-intensive than a comparable image. Unlike a text response, where a bad output can be discarded in milliseconds, a failed video generation costs real money and real time, and getting consistency across frames has been an industry-wide challenge that requires repeated iteration.
2. Pricing video is tricky: OpenAI priced Sora like it does chat, as a part of a flat $200 per month plan. For a serious video creator generating multiple clips daily, this is a bargain that OpenAI probably could not sustain. Flat subscriptions work well for text and code generation, where the marginal cost per query is low and usage is frequent. They work poorly for high-compute outputs where a small number of heavy users can consume a disproportionate share of resources.
The business logic of offering video inside a flat plan was questionable from the start, but it was probably necessary since Sora was a late entrant in video, and the download decline data suggests that even at that price, consumer demand did not materialise in volume. RunwayML (pricing) and Adobe’s Firefly have been building their pricing models around that reality, giving a limited number of credits for the plans that users buy, not all-you-can-eat pricing.
3. Why OpenAI chose coding over creation: The framing from OpenAI’s own leadership is instructive. Fidji Simo, the company’s applications chief, told employees in a memo quoted by the Wall Street Journal that the company “”cannot miss this moment because we are distracted by side quests”, and laid out a vision focused on “productivity in general and particularly productivity on the business front.” It’s a bit strange to call Sora a side-quest, if you ask me.
The WSJ also reported that OpenAI is now combining its ChatGPT desktop app, its coding tool Codex, and its browser into one “superapp.” The coding market is where the pressure is most acute. Claude has built a strong reputation among developers for coding tasks, and demand for it has been high enough to strain Anthropic’s own capacity. That people (including me) are frustrated with Claude’s regular timeouts indicates that coding has higher demand and lower supply. OpenAI has a low-hanging fruit right there, initially as a challenger to Claude, and potentially replacing it as the market leader.
Separately, Google’s video and image generation has improved significantly, especially with Nano Banana, and it has distribution advantages in that space that OpenAI does not have. Trying to fight Google on video diffusion while also catching up to Claude on coding was not a viable two-front strategy. Google has deeper pockets and several revenue streams that OpenAI doesn’t.
Reasoned is where I write about how AI is changing the world, whether it’s Commerce, Social Media, Content, Classifieds, Payments or even war. I publish twice (sometimes thrice) a week.
Do consider subscribing.
4. OpenAI was testing too many revenue streams at once: The hurried launch of apps, ecommerce and advertising, and the rollback of Ecommerce suggest that OpenAI was doing too many things at once, trying to test for what brings in revenue. Trying to be everything to everyone.
The Sora shutdown is the latest in a series of retreats. Earlier this month, OpenAI also scaled back its e-commerce plans (my take), having initially explored the possibility of enabling purchases directly inside ChatGPT before pulling back to a model focused on brand-owned ChatGPT apps. Advertising hasn’t fully rolled out either, and Claude’s anti-advertising ad during the Superbowl probably hit OpenAI’s positioning. Advertising has its own challenges (my take).
OpenAI reportedly aims for an IPO as soon as the fourth quarter of this year, and it needs to demonstrate both usage and margins. OpenAI needs revenue to raise the next round of funding, probably on before their IPO wherein they need to show margins and better revenue streams.
5. What the Disney deal really meant: The collapse of the Disney deal is the most visible collateral damage from the shutdown. Disney had agreed to a $1 billion equity investment in OpenAI as part of a three-year agreement that would have licensed more than 200 characters, including Luke Skywalker, the Toy Story cast, and the enormous (and enormously successful) Marvel roster of characters, for user-generated video inside Sora and ChatGPT. Disney told Variety it “respects OpenAI’s decision,” but a $1 billion investment is no longer proceeding.
The Disney deal collapse means more than money, because OpenAI was demonstrating that IP owners could work with it, and yet retain control. That proof of concept is now gone, as is any partnership with a major studio.
6. Why video is a regulatory nightmare: Video generation also carries policy and regulatory complexity that text does not: deepfake regulation, CSAM concerns, copyright liability, and an increasingly active global regulatory response. These are fragmented regulations globally, and tricker to execute in case of video. That adds to the cost and complexity of operating in the space. Running a consumer video platform means managing a content moderation challenge that is meaningfully harder than managing a text or image platform. It’s a pain to manage from a policy perspective.
India’s Synthetically Generated Information rules are a nightmare and probably unenforceable, but they’re still there. They shouldn’t be, but that’s another story.
7. Will OpenAI return to video eventually? Sam Altman has said the Sora team will now focus on world simulation research for robotics. That is a coherent longer-term use case: training robotic systems requires the ability to generate realistic physical environments at scale and understand how humans interact with them. The video model itself is not being deleted, and I wonder whether the physics of real-world interaction will be used to return to video in the future. Games do physics well, and Sora was using the Unreal Engine, as many games do.
8. Where will the users go? For the users who were generating content on Sora, the immediate question is where they go. RunwayML and Adobe Firefly are the most obvious alternatives, and both are better positioned to serve professional video creators with more expensive pricing models. Advanced users who were relying on Sora’s API will also need to migrate. These significant disruptions, still. Hobbyists will still use Nano Banana to do random things.