OpenAI Is Shutting Down Sora. It Burned $1 Million a Day and Lost Half Its Users.

OpenAI is shutting down Sora, the video generation platform it launched to widespread fanfare in late 2024. The app will close in April 2026, with the developer API following in September 2026, according to reporting from The Decoder citing a Wall Street Journal account of internal documents.

The numbers behind the decision are stark. Sora burned approximately $1 million per day in compute costs at its peak. Over the same period, its active user base dropped from roughly one million to around 500,000 — a 50 percent decline with no recovery trend. OpenAI's internal assessment reportedly concluded there was no viable path to making the product economically sustainable at its current capability level and cost structure.

What Killed Sora

The shutdown reflects a convergence of problems that were individually manageable but together proved fatal. Copyright exposure from AI-generated video content created ongoing legal risk. Low-quality deepfake videos produced by users generated reputational damage that was difficult to contain. And the fundamental unit economics of video generation — compute-intensive model inference against a largely free-tier user base — never closed.

Perhaps most telling: OpenAI canceled new training runs for next-generation video models entirely, rather than simply deferring them. The decision to stop investing in the next version of Sora before shutting the current version suggests the product team concluded that a more capable Sora would still not be commercially viable. The competitive dynamics also played a role — Anthropic's accelerating momentum in coding and enterprise forced OpenAI to make hard resource allocation decisions, and video generation lost.

Where the Team Goes Next

The Sora team will not disperse. OpenAI is redirecting their work toward world models for robotics — AI systems that build coherent representations of physical space and causality to enable robotic action. This pivot signals how OpenAI categorizes the underlying capability: Sora's video generation was always partly a step toward physical world modeling, and the robotics application is a higher-value path to commercializing that research than consumer video.

OpenAI is reallocating the freed compute capacity toward three priority areas: coding (where Claude Code and competitors have created intense pressure), enterprise applications, and agentic products. The reallocation reflects the company's judgment about where AI is generating durable revenue growth in 2026.

The Broader Lesson

Sora's shutdown is the most significant AI product failure from a frontier lab to date, measured by initial expectation versus outcome. It arrived with extraordinary press coverage, a waitlist, and genuine technical achievement. Its closure less than two years later reveals the difficulty of translating generative video capability into a sustainable consumer product — particularly when the cost base is high, the copyright environment is hostile, and the content moderation burden is severe. The labs building video generation successors are now working with a clearer picture of what that challenge actually requires.