OpenAI kills Sora

SORA LAUNCHED in September 2025 to the kind of breathless reception that Silicon Valley reserves for its most photogenic creations. OpenAI's AI video generator briefly hit number one in the Apple App Store, racked up four million downloads within a month, and landed a $1 billion licensing deal with Disney — the first time the Mouse House had ever handed its intellectual property to an AI platform. By December, more than 200 animated characters from Disney, Marvel, Pixar, and Star Wars were available for users to drop into AI-generated clips. Bob Iger called it "breathtaking growth." Sam Altman called Disney "the global gold standard for storytelling." Hollywood appeared to be embracing the machine. It was, by any measure, a spectacular debut. It lasted roughly six months.

On Tuesday, Altman told staff that the company would wind down the Sora consumer app, discontinue its developer API, and drop video functionality from ChatGPT entirely. The decision is part of a broader consolidation ahead of a potential IPO as early as the fourth quarter of 2026, in which OpenAI is merging ChatGPT, its Codex coding platform, and the Atlas browser into a single desktop "superapp" aimed squarely at enterprise productivity. Fidji Simo, OpenAI's chief of applications — hired from Instacart in May 2025 to impose product discipline on what had become a sprawling portfolio — framed the retreat with characteristic startup candor: the company could not afford to be distracted by "side quests" while Anthropic was eating its lunch in the enterprise market. According to an Axios analysis, Anthropic now captures roughly 73% of first-time enterprise AI spending. Claude overtook ChatGPT as the most-downloaded app in the United States in March 2026. Side quest, indeed.

The Sora shutdown is not an isolated product decision. It is the culmination of a strategic reckoning that has been building inside OpenAI for months. The company reported 900 million weekly active users as of February 2026 — more than double its figure from a year earlier — but that colossal reach was spread across too many parallel tracks, none of which were executing at the quality level leadership wanted. Simo's internal memo was blunt: "We realized we were spreading our efforts across too many apps and stacks, and that we need to simplify." In OpenAI's taxonomy, Sora was not killed for being bad. It was killed for being expensive and irrelevant to the business that will actually go public.

But Sora's demise is more than an internal housekeeping decision at one company. It is the sharpest signal yet that AI-generated video — the technology that was supposed to remake Hollywood by the end of the decade — has collided with the twin constraints of economics and indifference. The economics are punishing. Estimates compiled by Forbes and analyst Deepak Mathivanan of Cantor Fitzgerald suggest OpenAI was spending approximately $15 million per day to power Sora, annualizing to more than $5.4 billion. Each ten-second clip cost around $1.30 in raw GPU time, requiring roughly 40 minutes of total GPU computation — or eight to ten minutes split across four GPUs running in parallel. Assuming even a quarter of Sora's 4.5 million users were active and posting an average of ten videos daily, the platform was churning out more than 11 million videos a day. Altman himself conceded that users were mostly "making funny memes to send to their three friends" — a use case for which, as he put it, no advertising model can cover the cost.

Video generation is the most resource-intensive task in consumer AI, and OpenAI was effectively subsidizing a social media platform that generated engagement but not revenue. When free credits ran out, the company experimented with selling ten additional video generations for $4 — a price point that revealed the gap between what users would pay and what the compute actually costs. At the API level, Sora 2 runs $0.10 to $0.50 per second of generated video depending on resolution and model tier. A single polished ten-second clip can cost anywhere from $1 to $5 before factoring in the multiple iterations most creators need. By January 2026, OpenAI had already discontinued the free tier entirely, requiring at minimum a $20-per-month ChatGPT Plus subscription for access. The market's verdict was quiet but decisive: usage flatlined.

The indifference, meanwhile, extends well beyond OpenAI's user base. Deloitte predicted that major film and television studios would allocate less than 3% of their production budgets to generative AI tools in 2025, while an Atlassian survey found that 96% of CEOs reported no meaningful return on investment from AI initiatives broadly. Hollywood has proven eager to experiment with AI for back-office functions — contract management, localization, pre-visualization, script breakdowns for shooting schedules — but singularly reluctant to put generative video anywhere near a camera. Guillermo del Toro said he would rather die than use the technology in his films (a sentiment that, albeit dramatic, captures the prevailing mood among A-list directors). SAG-AFTRA contracts expire again in 2026, and the prospect of another AI-fueled labor stoppage hangs over every studio lot in Burbank. The 2023 strikes, the longest in the union's history, were animated in no small part by fears over AI; the next round of negotiations promises to be equally combustible.

Yet the competitive field is hardly barren. Runway, Pika, Google's Veo 3.1, and the Chinese upstart Kling continue to iterate at speed. Runway's Gen-4 architecture remains the industry standard for professional creators who need frame-level control and integration with tools like Adobe Premiere. Veo 3.1 has set a new benchmark in native audio synchronization, delivering 4K output with dialogue and environmental sound generated alongside the visuals. Pika has carved a niche in fast, stylized social content — its viral Pikaffects tool generated a single video with more than 19 million views on TikTok. Kling, meanwhile, has emerged as a formidable dark horse from China, offering two-minute videos with camera controls at prices that undercut Western competitors by a wide margin. Average cost per minute of AI-generated video dropped 65% between 2024 and 2025, according to industry estimates, and inference costs are projected to fall fivefold again by 2027. The tools are getting cheaper, better, and more numerous. What they are not getting, at least not yet, is indispensable.

The core problem is structural. Today's diffusion models treat each generation as statistically independent — there is no persistent character state, no memory across shots, no reliable physics over extended durations. OpenAI itself has acknowledged that even Sora's deployed version suffers from "many limitations," including unrealistic physics and an inability to handle complex actions over longer clips. A creator can produce a stunning five-second clip; stitching twenty of them into a coherent scene remains, as one independent filmmaker discovered after a fourteen-hour detour through DaVinci Resolve, an exercise in brute-force patience. For social media, where attention spans are measured in heartbeats and visual inconsistency is a feature rather than a bug, this is tolerable. For Hollywood, where a single continuity error can halt a $200 million production, it is a dealbreaker. The gap between "impressive tech demo" and "reliable production tool" has narrowed, but it has not closed.

OpenAI's retreat puts the Disney partnership in a particularly awkward light. The three-year licensing deal, announced to enormous fanfare in December 2025, was designed to showcase Sora as the platform where fans could create short-form videos featuring Mickey Mouse, Darth Vader, and Iron Man. Disney was to curate the best user-generated content for Disney+, turning its streaming service into something resembling an AI-powered TikTok for intellectual property. A joint steering committee was established to monitor creations against a voluminous brand appendix. The deal was exclusive, at least initially. Iger hinted at exclusivity "at the beginning of the three-year agreement" without specifying its duration. Copyright experts hailed the arrangement as a template for the future — a negotiated partition between AI and content that could end the cold war over training data and likeness rights.

Three months later, the specific platform the deal was built around is being mothballed. Disney retains its equity stake and its broader API relationship with OpenAI — the company is still deploying ChatGPT for its employees and using OpenAI's models to build tools for Disney+ — but the consumer-facing Sora product that was supposed to put Iron Man in the hands of every teenager with a prompt is going dark. Whether the character licensing migrates into some future iteration of ChatGPT's image and video capabilities remains unclear. What is clear is that the deal's splashiest selling point — Sora as the home of Disney-powered AI video — has evaporated before the first licensed videos were widely available.

The broader lesson is one the AI industry has been slow to absorb. Generative video is extraordinarily good at producing wonder and extraordinarily bad at producing revenue. The demos are cinematic; the unit economics are brutal; and the professional use cases, while real, remain confined to pre-production, B-roll, and social media — segments that are lucrative for niche startups like Runway but nowhere near large enough to justify the compute expenditure of a company burning $15 million a day. OpenAI's conclusion, implicit in its actions if not in its words, is that coding agents and enterprise productivity tools are demonstrably closer to product-market fit than anything involving pixels in motion.

The AI video revolution has not been cancelled. It has been repriced. The question is no longer whether machines can generate cinema — the demos have settled that — but whether anyone can build a business around it before the next model generation renders today's economics obsolete. For now, the most advanced AI company on Earth has looked at its own video product and decided the pixels were not worth the watts. ■

For more, join 75,000 subscribers getting tech's favorite brief here