AI Video Got Good Enough to Worry Hollywood. OpenAI Killed Its Flagship App Anyway.

Sora, OpenAI's text-to-video model, generated some of the most convincing AI video the public had seen when it launched for ChatGPT Plus and Pro subscribers in December 2024 — photorealistic footage with objects and scenes that held together across frames in a way earlier video models couldn't manage. By April 2026, OpenAI shut the consumer product down entirely. The story of why is a more useful read on where AI video actually stands than the demo reels were.

What actually got shut down, and when

OpenAI notified developers on March 24, 2026 that the Sora app and web experience would close, shut it down on April 26, 2026, and scheduled the separate Sora 2 API for removal by September 24, 2026. That's a full wind-down of the consumer product, not a pause or a rebrand — a notable reversal for a tool that had generated some of the most viral AI demos of the previous two years.

Why: the unit economics never worked

Video generation is enormously more computationally expensive than text or image generation, and usage numbers reportedly never hit the projections that would have justified that cost at consumer pricing. OpenAI redirected the underlying resources toward reasoning models, coding tools, and core ChatGPT improvements — areas more directly tied to the enterprise revenue the company has prioritized. The gap between an impressive demo reel and a product people would pay enough to sustain turned out to be wide: output quality was inconsistent between generations, rendering was slow, and usage caps discouraged the kind of iterative, try-it-ten-times workflow that creative tools usually need to be genuinely useful.

The core problem wasn't that Sora's video looked bad — it was that making it look good, consistently, at a price OpenAI could sustain, turned out to be a much harder problem than making it look good once, for a demo.

What happened with Hollywood

Hollywood did try. In December 2025, Disney announced a three-year licensing deal that would have brought more than 200 of its characters to Sora, alongside a planned $1 billion equity investment in OpenAI. That deal ended in March 2026, reportedly before any money changed hands, when OpenAI decided to exit video generation. Even before that, three practical obstacles limited Sora's use in real production. Sora couldn't reliably produce the same character consistently across multiple clips, which is a baseline requirement for any actual production. Post-2023 strike agreements between studios and unions placed real restrictions on how AI tools could be used in professional production. And professional filmmaking demands a level of precise, repeatable control that text-prompt-based generation, as of Sora's design, couldn't reliably provide — a director can't easily tell a text prompt to match a previous shot's exact lighting and camera angle.

Google's Veo is playing a different game

Google's Veo 2 and Veo 3, the latter with native audio generation, are generally regarded as the most credible enterprise-facing alternative to what Sora attempted, partly because Google can distribute and monetize video generation through YouTube's existing creator and advertising infrastructure rather than needing a standalone consumer app to carry its own economics. That's a structurally different bet than OpenAI's: subsidize video generation as a feature of a larger platform with other revenue streams, rather than sell it as the product itself.

What this actually says about AI video's trajectory

Sora's shutdown isn't evidence that AI video generation is a dead end — the underlying model capability was real, and demand for AI-assisted video tools in advertising, marketing, and lower-stakes content production continues to grow. It's evidence that "impressive demo" and "sustainable product" are different bars, and that video generation's computational cost makes clearing the second bar much harder than it's been for AI's text and image tools. The more durable near-term use cases — ad creative, social content, previsualization — are quieter and less viral than "AI can make a movie," which may be exactly why they're the ones actually shipping.