OpenAI introduces Sora 2, its new AI video model

OpenAI has introduced Sora 2, an artificial intelligence model that generates videos with synchronized sound, dialogue, and more realistic physics. The company also launched a social app for creating, remixing, and sharing these scenes with friends.

Videos that follow the rules of the real world more closely

The main difference from earlier versions is how the model represents motion and mistakes. If a basketball player misses a shot, for example, the ball might bounce off the backboard instead of suddenly appearing inside the hoop.

OpenAI says Sora 2 can generate scenes that are difficult for other systems, such as Olympic gymnastics routines, a somersault over a paddleboard, or a triple axel with a cat clinging to the skater. The model still makes mistakes, but it tries to maintain object continuity and follow physics more closely.

That matters because a video AI does not just need to produce attractive images. It also needs to understand what happens when something goes wrong. In a sports scene, a fall or a missed throw can matter as much as a perfect movement.

The model also follows more complex instructions across multiple shots. It can maintain the appearance of a person, object, or setting while changing the framing, style, or action. OpenAI says it works especially well with realistic, cinematic, and anime-style videos.

Dialogue, effects, and people inside the scene

Sora 2 generates video and audio at the same time. That includes conversations, ambient sounds, and effects such as impacts, footsteps, or a bouncing ball. In practice, you can describe a scene and receive a more complete sequence without having to add each sound separately afterward.

One of its most notable features is called cameos. After a short video and audio recording to verify your identity, you can appear in AI-generated scenes with a representation of your appearance and voice.

The feature can also be used with friends, animals, or objects. Each person decides who can use their cameo, can withdraw that permission, and can delete any video that includes it. Videos featuring their cameo, including those created by other people as drafts, remain available for them to review.

Sora becomes a social network

OpenAI has launched an iOS app called Sora. Inside it, you can:

• Create videos from prompts.
• Remix videos from other users.
• Discover content in a customizable feed.
• Invite friends to appear in your scenes.

The launch begins on September 30, 2025 in the United States and Canada by invitation. The company plans to expand access to other countries and also offer Sora 2 at sora.com after you receive an invitation.

Initial use will be free, although it will be limited by available server capacity. OpenAI is also preparing Sora 2 Pro, an experimental higher-quality version for ChatGPT Pro users, and plans to release the model through an API so other applications can integrate it.

The challenge: creating without turning it into another endless feed

OpenAI says it does not want to optimize the app so you spend as much time as possible scrolling through videos. The feed will prioritize people you follow or interact with, along with content that may inspire you to create something new.

The company also promises controls for adjusting recommendations through natural language, regular reminders about user well-being, and daily limits for teenagers. Parents will be able to manage those limits, turn off personalization, and control private messages from ChatGPT.

The system will have to deal with familiar problems in video generation: impersonation, non-consensual use of someone's likeness, harassment, and the creation of harmful content. OpenAI says it will combine automated filters with human moderators, especially when reviewing potential cases of bullying.

For you, the most visible change is that creating a video with your appearance, voice, and a coherent scene will become much easier. The deeper change lies elsewhere: OpenAI is using these models as a step toward systems capable of simulating the physical world. Sora 2 still fails, but the way it fails will matter as much as the quality of its images.