Sora

The name comes from the Japanese word for sky — chosen, its creators said, because it "evokes the idea of limitless creative potential." That framing was not marketing copy. When OpenAI demonstrated Sora on February 15, 2024 — videos of woolly mammoths trotting through snow, a Tokyo street scene shot by a sweeping virtual camera, a woman walking through rain-slicked city streets — the reaction across the creative and technology industries was closer to vertigo than excitement. The videos were not almost-good. They were indistinguishable from Hollywood footage to most viewers. Film directors, cinematographers, and VFX supervisors who had spent decades learning their craft watched the demos and went quiet.

The architecture behind the illusion was as surprising as the output. Sora is not a conventional diffusion model. Its co-leads — Tim Brooks and William (Bill) Peebles, both PhD graduates of Berkeley's AI Research Lab under Alyosha Efros — built it around the Diffusion Transformer (DiT) architecture that Peebles had proposed in a 2022 paper that CVPR initially rejected for "lack of innovation." The irony that the foundational paper of one of the most consequential AI products in history was initially dismissed by a peer review committee is now well-documented. DiT treats video not as a sequence of frames but as a field of spacetime patches — visual tokens analogous to the text tokens in a language model — which allows Sora to reason about scenes, physics, and temporal coherence at a structural level rather than pixel by pixel. The result is a model that, as Peebles noted, automatically generates different camera angles without being told to, and that Tim Brooks described as having "learned how to create 3D graphics from its dataset alone." The sky, it turns out, was not the limit. It was the floor.

Sora 2, launched September 30, 2025 alongside a dedicated iOS app, extended the original's capabilities significantly: videos up to 25 seconds, native synchronized audio including human dialogue with accurate lip-sync, ambient environmental sound, and composed background audio — all generated simultaneously with the video in a single pass. The app introduced Cameo, letting users insert themselves or others into AI-generated scenes; Blend, for combining clips; and Storyboard, for assembling sequences. The New York Times called it "a social network in disguise" — a product that Meta and TikTok had sought to build and couldn't. The copyright chapter has been thorny: Japan's Content Overseas Distribution Association demanded OpenAI stop training on Ghibli and Square Enix content; families of Robin Williams and George Carlin urged action against AI-generated deepfakes; OpenAI restricted specific celebrity likenesses and began giving rights holders opt-out controls. OpenAI's response has been iterative rather than categorical — a reflection of the tension between a product whose creative potential is maximal precisely because of its lack of restrictions, and a company that has staked its commercial position on being the responsible actor in frontier AI.