OpenAI previews Sora, a text-to-video model

OpenAI published sample output from Sora, a diffusion transformer that generated video up to a minute long from text prompts. The clips — a woman walking through neon-lit Tokyo, mammoths crossing a snowy plain, a paper aeroplane over a forest — held their subjects consistent across shots and camera moves in a way previous video systems had not.

The technical note framed the system as more than a video generator. Sora operated on spacetime patches, an approach analogous to tokens in language models, and OpenAI argued that scaling it produced emergent simulation of physical consistency, object permanence and three-dimensional coherence — proposing video generation as a path to models that understand the physical world. The paper’s title, “Video generation models as world simulators,” was the claim.

The limitations were shown alongside: physics that failed under stress, objects appearing and vanishing, hands and glass behaving impossibly. Critics noted that the released examples were curated and that no quantitative evaluation accompanied them.

Sora was not released publicly at announcement. It went to red-teamers and a small group of visual artists, with general availability arriving in December 2024 — a ten-month gap that OpenAI attributed to safety work around likeness, misinformation and election integrity, and that competitors used. Runway, Luma, Google’s Veo, Kuaishou’s Kling and Chinese open-weight video models all shipped in the interval.

The announcement’s effect on the film and advertising industries was immediate and largely apprehensive, arriving a few months after the 2023 Hollywood strikes had settled AI provisions that assumed slower progress. It also moved deepfake concern from a research topic to a production one, in a year in which more than half the world’s population lived in countries holding elections.