OpenAI Killed Sora. The Unit Economics Were Never Going to Work.

OpenAI announced on March 24, 2026 that it was discontinuing the Sora app, the Sora API, and Sora.com. The app shut down on April 26 and the API followed on September 24. The company cited a strategic refocus on physical AI and foundation models for robotics. The proximate economic cause, reported by multiple outlets citing OpenAI internal communications, was a cost per generated minute that could not be recovered at any price point consumers demonstrated willingness to pay. Disney, OpenAI’s highest-profile Sora enterprise partner, did not proceed with its partnership and pivoted to Runway Gen-4 for professional video workflows.

The Unit Economics That Killed Sora

Two different cost figures circulate for Sora, and they measure different things. The Wall Street Journal reported losses of roughly $1 million per day. A separate third-party estimate put inference costs at approximately $15 million per day at peak usage, which is a spike figure rather than a sustained daily average. Our fuller account of the shutdown reconciles both. The distinction matters, because the two numbers differ by more than an order of magnitude and only one of them describes the ongoing burn.

The per-generation economics are clearer than the daily totals. Generating a single minute of video was estimated to cost OpenAI between $5 and $15 depending on resolution and complexity. A user generating 10 videos per day, not unusual for content creators experimenting with the tool, could cost tens to low hundreds of dollars per day. OpenAI’s Pro subscription costs $200 per month. On those estimates, a single active Sora user could consume more than a month’s subscription value in a day of heavy generation.

Set that against Appfigures’ estimate of roughly $2.1 million in lifetime in-app revenue and the shape of the problem is legible regardless of which daily figure you accept. On the Journal’s numbers, the six months between the September 2025 app launch and the March 2026 shutdown burned roughly $180 million against that $2.1 million. Video generation is compute-intensive in a way that text generation is not. An LLM generates a text response in milliseconds using a few cents of GPU compute. A video model generates a few seconds of video using minutes of GPU compute costing dollars. That gap, not any single headline figure, is why the pricing model that works for ChatGPT could not work for video.

The Disney Deal and What Killed It

The $1 billion Disney partnership, announced in late 2025, was supposed to be Sora’s path to sustainability: enterprise licensing at prices that could cover inference costs. Disney planned to use Sora for pre-visualization, concept art animation, and rapid prototyping of visual effects. Reporting on the collapse indicates Disney’s creative teams found Sora’s output insufficient for professional production workflows. Generated videos lacked temporal consistency (objects changed between frames), physical accuracy (gravity, lighting, and material properties were unreliable), and creative controllability (directors could not reliably specify camera angles, character positioning, or scene composition). Disney has not published its own account of the decision.

The collapse exposed a gap between demo quality and production quality. Sora’s demo videos were selected from many generations. Professional production requires reliable, repeatable output on the first or second attempt, not the best result from fifty.

Why Frontier Video Generation Economics Do Not Work Yet

Compute cost per minute of frontier-quality video is estimated at $2 to $8, a reported range rather than a disclosed figure. Sora’s consumer access came through the $20 per month ChatGPT Plus bundle. At that price, on those estimates, only a few minutes of video per user per month are sustainable before gross margin goes negative. Observed willingness to pay for video-only subscriptions sits around $10 to $15 per month based on Runway and Pika pricing.

Runway, Pika, and Kling have survived and grown in the consumer video market by running more efficient models at lower quality points with better pricing structures. The lesson from Sora is not that AI video is economically impossible. It is that frontier-quality AI video at consumer price points was not economically viable at 2026 compute costs. Runway Gen-4 is not as technically impressive as Sora was. It is economically sustainable. For a product, that is the metric that decides survival.

The Pivot to Robotics Simulation

OpenAI is redirecting the Sora team toward world simulation for robotics applications. This pivot makes technical sense: video generation models build internal representations of physical world dynamics, including how objects move, how light behaves, and how materials interact. Those representations, even if insufficient for Hollywood-quality video, may be sufficient for training robotic systems that need to predict how the physical world will respond to their actions.

The robotics application sidesteps Sora’s consumer product failures. Robotics training does not require aesthetic quality, it needs physical accuracy. It does not require real-time generation, it can use batch processing. It does not require consumer pricing, industrial customers pay enterprise rates. Whether Sora’s physics representations are accurate enough for robotics training is an open empirical question that requires testing against real-world robotic performance, and no such results have been published.

The Robotics Pivot Explains the Compute Redirect

OpenAI’s physical AI roadmap, which the company has been building toward since its Figure partnership and internal robotics research from 2025, requires substantial compute for training foundation models on embodied agent data: video of physical manipulation, sensor data from robotic arms, and simulation runs. The compute previously allocated to Sora inference is being redeployed onto robotics foundation model training, which OpenAI has positioned as a higher-priority path ahead of a possible IPO.

For that timeline, killing Sora removes a cost center that was generating negligible revenue. The move improves unit economics immediately and repositions the technology as an enterprise research tool rather than a consumer product with negative margins. The narrative shifts from a money-losing video product to physics simulation technology for robotics. The underlying technology is largely the same. The framing is the difference, and framing is what an IPO prospectus sells.

Sora’s run is a data point in a larger pattern: the hardest AI product challenge is not building frontier capability but monetizing it at the cost structure required to sustain it. OpenAI remains among the strongest model builders in the industry. The Sora shutdown is a product strategy decision, not a research failure. Whether the robotics bet pays off is the question that will determine whether the compute was well spent.

Sources: OpenAI’s Sora discontinuation announcement; Wall Street Journal reporting on daily losses and the Altman decision; Appfigures lifetime revenue estimate; Disney partnership reporting; Runway Gen-4 launch announcements.

Updated 2026-08-23: corrected the discontinuation date to March 24, 2026 and the app launch to September 2025, reconciling this piece with our three companion Sora articles; attributed the $1 million per day loss figure to Wall Street Journal reporting and reclassified the $15 million per day figure as a third-party peak-inference estimate rather than a sustained cost; removed an extrapolation of that peak figure across months; marked per-minute cost ranges as estimates; and noted that Disney has not published its own account.