Sora 2

Sora 2 is OpenAI's flagship video and audio generation model, announced on 30 September 2025. According to OpenAI, Sora 1 (February 2024) was the "GPT-1 moment" for video — Sora 2 is described as the "GPT-3.5 moment", with substantially improved physics, realistic modelling of failure (for example a basketball rebounding off the backboard rather than teleporting to the hoop), and better controllability of multi-shot instructions.

Architecture and audio

Sora 2 is the successor to Sora 1 — a diffusion transformer (DiT) trained in latent space. Unlike its predecessor, the model is a general-purpose video and audio generation system: it produces sophisticated background soundscapes, dialogue and sound effects that are synchronised with the visuals.

Characters feature

Sora 2 introduces a "characters" feature — after a short one-time video-and-audio recording that verifies the user's identity, the model can insert that person (or an animal or object) into any generated scene with high fidelity in appearance and voice. The user retains full control over their likeness: they decide who can use it, can revoke access at any time and can view every video that contains their character, including drafts created by other people.

Sora app

Together with the model, OpenAI released a social iOS app called "Sora", powered by Sora 2. In the app, users can create and remix generations, discover videos in a customisable feed and insert themselves or their friends via the characters feature. Initial rollout: U.S. and Canada, invite-based. Once invited, access was also available through sora.com. ChatGPT Pro subscribers received access to an experimental higher-quality Sora 2 Pro variant.

Status

According to OpenAI, the Sora product (the social app) was discontinued on 26 April 2026. OpenAI had previously announced plans to release Sora 2 in the API as well. Sora 1 Turbo remained available alongside Sora 2.