OpenAI launches Sora 2 and an AI-generated video social network

OpenAI today unveiled Sora 2, a new version of its video generator that can produce scenes with dialogue, sound effects, and more coherent movement. The model arrives alongside Sora, a social iPhone app where users can create videos, appear in them through their digital likeness, and publish them to a TikTok-style feed.

The launch broadens the product’s scope. The original Sora was primarily an audiovisual creation tool; its successor also aims to power a social network where virtually all content is made with artificial intelligence.

Sora 2 adds sound and improves how objects behave

Sora 2 generates video and audio together, including dialogue synchronized with characters and ambient sounds. That is a significant step up from the first video models, which typically produced silent clips and required users to add voices, music, or effects later with other tools.

OpenAI also says it has improved physics simulation. That does not mean the model understands the laws of physics like a scientific simulator; rather, it does a better job of preserving objects’ trajectories and the consequences of an action.

The company explains it with a simple example: earlier models could warp a scene to fulfill an instruction, even making a missed basketball shot appear inside the hoop. OpenAI wrote in a blog post: “Prior video models are overoptimistic — they will morph objects and deform reality to successfully execute upon a text prompt. For example, if a basketball player misses a shot, the ball may spontaneously teleport to the hoop. In Sora 2, if a basketball player misses a shot, it will rebound off the backboard.” Demonstrations released by OpenAI include beach volleyball, gymnastics, diving-board jumps, and skateboard tricks — activities that are especially difficult because any movement error is immediately apparent.

These are examples selected by the company itself, so it remains to be seen how the model performs with less controlled prompts, longer scenes, or interactions involving multiple objects. Plausible physics can reduce conspicuous errors, but it does not make every generated video a reliable depiction of reality.

“Cameos” turn identity into raw material

The app’s central feature is called cameos. Users record their appearance and voice once to verify their identity and create a representation that they can later insert into generated scenes. They can also authorize friends to use that cameo, including in videos featuring multiple people.

Permission can be revoked, but the system opens up a sensitive area. Authorizing someone to use your likeness does not guarantee that you approve every specific scene. A trusted contact could create a humiliating prank, a fabricated statement, or a video that ends up circulating beyond Sora.

The difference from traditional manipulated videos is the scale: users no longer need video-editing skills or a dedicated model trained on numerous photos. The platform itself brings identity capture, generation, and distribution together.

A personalized feed powered by Sora and ChatGPT data

Videos can be published to a vertical feed governed by algorithmic recommendations. To rank them, OpenAI will consider activity within Sora, approximate location obtained through the user’s IP address, engagement with previous posts, and — if the user allows it — their ChatGPT conversation history. The last data source can be disabled.

The connection gives OpenAI an advantage over a social network starting from scratch: ChatGPT already knows the interests that millions of users have expressed directly. It also raises an unusual privacy question. A conversation with an assistant about studying, work, or personal advice could end up influencing the entertainment shown by another app from the same company.

Sora includes parental controls managed through ChatGPT. Adults will be able to adjust infinite-scroll limits, turn off personalization, and decide who can send direct messages to a minor. Their effectiveness will depend both on the platform’s design and on whether families configure and understand them.

Limited launch and free generation at first

The app is initially available on iOS, in the United States and Canada, with access by invitation. OpenAI expects to expand it to other countries, while ChatGPT Pro subscribers will be able to try Sora 2 Pro without needing an invitation to the social network.

Use will be free at launch. The first planned monetization mechanism is to charge for additional generations when demand is high. This approach lowers the barrier to entry, although generating video consumes far more computing resources than producing text or images.

Deepfakes, copyrighted characters, and moderation

Sora brings together two problems that could previously be addressed separately: the creation of synthetic material and its algorithmic distribution. A deceptive video no longer has to leave the tool to find an audience; it can be generated and circulated within the same service.

Cameos introduce an explicit consent system for identity, but they do not eliminate the risk of abusive use or resolve what happens when a generated recording is downloaded and reposted. Moderation will have to distinguish parody, harassment, impersonation, and political manipulation without relying solely on victims to discover the content and report it.

Intellectual property is another unresolved conflict. Video models can recreate recognizable characters, visual styles, and fictional universes, but the technical ease of doing so does not grant rights to those works. For studios, artists, and franchise owners, the challenge will be preventing a social platform from filling up with commercial imitations or unauthorized scenes before claims can be processed.

OpenAI is not merely introducing a more capable model. It is testing whether synthetic video can become a mass-market form of consumption. International expansion, the first moderation decisions, and effective control over personal likenesses will determine whether Sora works as a creative tool or multiplies the problems already plaguing video-based social networks.