OpenAI has released Sora 2, its flagship video and audio generation model.
The original Sora model from February 2024 was widely regarded as the GPT‑1 moment for video—the first time video generation began to show realistic behaviours such as object permanence as a result of scaled-up pre-training compute. Since then, the Sora team has focused on training models with more advanced world simulation capabilities, believing these systems are essential for training AI that deeply understands the physical world. A key milestone in this progress is mastering both pre-training and post-training on large-scale video data, a field still in its early stages compared to language modelling.
Sora 2 represents what may be the GPT‑3.5 moment for video. The model is capable of generating complex actions such as Olympic gymnastics routines, backflips on paddleboards with realistic physics, and triple axels involving cats—all with a level of realism previously unseen in video generation systems.
Whereas prior video models often deform reality or bend the laws of physics to fulfil text prompts (e.g., teleporting basketballs into hoops), Sora 2 demonstrates greater fidelity to physical laws. In scenarios like a missed basketball shot, the ball rebounds naturally, rather than behaving unrealistically. Importantly, the model appears to simulate internal agents, making errors that resemble human or character-like decisions, further enhancing realism. This capacity to model failure, not just success, is critical for developing reliable world simulators.
Sora 2 is also a significant advancement in controllability. It can follow intricate, multi-shot instructions while accurately maintaining world state, and it excels across cinematic, realistic, and anime visual styles. As a general-purpose video and audio generation system, it produces high-quality background soundscapes, speech, and sound effects.
The model allows real-world elements to be directly injected into generated content. For example, a video of a person can be used to place them accurately—with likeness and voice—into any Sora-generated environment. This applies broadly to people, animals, and objects.
Although Sora 2 is not without flaws, Open AI says it validates the premise that scaling neural networks on video data brings AI closer to simulating reality.
Deployment of Sora 2
OpenAI says it believes the journey toward general-purpose simulation and AI systems should be enjoyable for users. The Sora team began experimenting with the “upload yourself” feature months ago and found it to be a natural evolution in communication—progressing from text to emojis to voice notes to immersive video cameos.
Alongside the model, OpenAI has launched a new social iOS app called “Sora,” powered by Sora 2. The app enables users to create and remix content, discover new videos via a customised feed, and appear in each other’s creations through the “cameos” feature. This feature allows users to insert themselves into any Sora scene after a brief video and audio verification. Early internal testing at OpenAI has led to new social connections and enthusiastic feedback about the experience.
Launching Responsibly
Recognising potential issues such as doomscrolling, addiction, and algorithmic isolation, OpenAI says it has has taken steps to launch Sora responsibly. New recommender algorithms—guided by natural language and built with existing OpenAI models—give users control over their feed. The app periodically checks in with users’ wellbeing and allows them to adjust settings accordingly.
By default, the feed prioritises content from people users interact with and videos likely to inspire further creation, rather than maximising screen time. The app is designed for creation, not passive consumption.
“Cameos” are central to the Sora experience, offering a new and compelling form of communication. The app is launching on an invite-only basis to encourage users to join with friends. At a time when other platforms are de-emphasising social graphs, OpenAI believes cameos will strengthen community engagement.
Teen wellbeing is also a priority. The app limits the number of generations teens can view per day and imposes stricter cameo permissions. Safety mechanisms include automated moderation and dedicated teams to address bullying. Sora also launches with parental controls via ChatGPT, allowing guardians to manage infinite scroll limits, algorithm personalisation, and direct messaging settings.
Users have full control over their likeness with cameos, including the ability to revoke access or delete videos. All videos containing cameos—even drafts—are accessible to the cameo subject at any time.
OpenAI says it has addressed safety issues related to likeness use, content provenance, and harmful generation in its Sora 2 Safety documentation. Unlike many other platforms, OpenAI is not currently pursuing monetisation models that may compromise user wellbeing. Any future monetisation—such as allowing payment to generate additional content when compute is constrained—will be communicated transparently.
OpenAI says it sees Sora 2 as the start of a new era of co-creative digital experiences, with optimism that this platform will provide a healthier outlet for creativity and entertainment.
Sora 2 Availability
The Sora iOS app is now available for download. Users can sign up in-app for push notifications once access is granted. The initial rollout begins in the United States and Canada, with plans to expand to additional regions. Upon receiving an invite, users can also access Sora 2 via sora.com. The platform is initially free with generous usage limits, although these may be subject to compute availability.
ChatGPT Pro users will gain access to an experimental, higher-quality Sora 2 Pro model on sora.com and eventually within the iOS app. OpenAI also plans to integrate Sora 2 into its API. Sora 1 Turbo remains available, and users’ previous content will continue to reside in their sora.com libraries.
Visit https://openai.com/sora/
