About
Sora is a family of OpenAI generative models for video (and video with audio), built on the diffusion transformer (DiT) architecture. The models learn to represent video and images as collections of spacetime patches, analogously to tokens in large language models, and are trained in latent space (latent diffusion). The first model — Sora (Sora 1) — was unveiled on 15 February 2024 and publicly released on 9 December 2024 as Sora Turbo. Sora 2, announced on 30 September 2025, adds synchronised audio generation (dialogue and sound effects), substantially improved physics and controllability, and a "characters" feature (inserting verified likenesses). Together with Sora 2, OpenAI launched the social iOS Sora app; the consumer product was discontinued on 26 April 2026.