OpenAI Launches Sora 2 and a Social iOS App Built Around It

OpenAI released Sora 2, its flagship video and audio generation model, on September 30, alongside a new invite-based social iOS app called "Sora" that launches first in the U.S. and Canada.

The Sora team frames the release as a generational jump. The original Sora model from February 2024 was, in their words, "the GPT-1 moment for video" — the first sign that video generation was starting to work. With Sora 2, OpenAI says it is "jumping straight to what we think may be the GPT-3.5 moment for video."

The claimed capability jump is concrete. Sora 2 can generate Olympic gymnastics routines, backflips on a paddleboard that "accurately model the dynamics of buoyancy and rigidity," and a figure skater landing a triple axel with a cat on her head — sequences the company describes as "outright impossible" for prior video generation models.

The more technical claim concerns failure. Previous video models, OpenAI argues, are overoptimistic: if a basketball player misses a shot, the ball teleports into the hoop. In Sora 2, the ball rebounds off the backboard. "This is an extremely important capability for any useful world simulator — you must be able to model failure, not just success," the team writes. The model still makes mistakes, which OpenAI says often appear to stem from the internal agent the model implicitly simulates.

The stakes go beyond novelty clips. OpenAI argues that video models with advanced world simulation capabilities "will be critical for training AI models that deeply understand the physical world," and that general-purpose simulators and robotic agents "will fundamentally reshape society." Pre-training and post-training on large-scale video data remain, by the company's own admission, "in their infancy compared to language."

Sora 2 also generates synchronized audio — background soundscapes, speech, and sound effects — and handles multi-shot instructions while persisting world state across them. A "characters" feature lets users record a short one-time video-and-audio clip in the app to verify identity and capture their likeness, then insert themselves into any Sora-generated scene with accurate appearance and voice. OpenAI says the capability works for "any human, animal or object."

The app and the feed

The social app is the deployment vehicle. Users can create, remix each other's generations, and browse a customizable feed. OpenAI is positioning it as a deliberate counter to incumbent social platforms. The default feed is heavily biased toward people users follow, and the company states plainly: "We are not optimizing for time spent in feed." The recommender system, built on OpenAI's existing large language models, can be instructed through natural language, and the app periodically polls users on their wellbeing and offers feed adjustments.

Teen protections include default daily limits on feed generations, stricter character permissions, scaled-up human moderation for bullying cases, and parental controls via ChatGPT that let parents override scroll limits, turn off personalization, and manage direct messages. On monetization, OpenAI says its only current plan is to eventually let users pay to generate an extra video when demand exceeds available compute.

On likeness, the company says users control their character end-to-end: only the owner decides who can use it, access can be revoked, and any video containing the character — including other people's drafts — remains viewable by that person.

Availability

Sora 2 is free at launch with "generous limits" subject to compute constraints. ChatGPT Pro users get access to an experimental higher-quality model, Sora 2 Pro, on sora.com, with app access planned. OpenAI also plans an API release. Sora 1 Turbo remains available, and existing creations stay in users' sora.com libraries.

The invite-based rollout — "to make sure you come in with your friends" — signals OpenAI's bet that a social graph built around AI-generated characters, at a moment when "all major platforms are moving away from the social graph," is the differentiation that matters. Whether instructable feeds and teen limits hold up under real-world usage will determine if that bet is more than launch-week positioning.