OpenAI debuts Sora 2 AI video generator app with sound and self-insertion cameos, API coming soon

OpenAI today announced the release of Sora 2, its latest video generation model, which now includes AI generated audio matching the generated video, as well.
It is paired with the launch of a new iOS app simply called Sora, that allows users to insert and edit with AI videos of themselves and their friends alongside them with a new “Cameo” feature — though OpenAI says there are robust protections and identity safeguard measures in place to prevent someone’s identity from being inserted into AI videos without their consent or approval.
In addition, the company said a Sora application programming interface (API) is in the works that will allow third-party developers to pipe the new Sora 2 model into their own video editing applications, unlock new, more fine-grained and professional editing capabilities, and generally push the frontier video generation model into new directions.
Furthermore, OpenAI says an Android mobile app is also in the works. WIRED magazine (which my wife runs) first leaked the news that a Sora 2 AI mobile app was being developed about a day before the official OpenAI announcement.
ChatGPT users in the U.S. and Canada will be first to get their hands on the new model and app, though OpenAI said it plans to expand access to other countries in the coming days and weeks. Sora 2 is available free with usage limits for all users, while ChatGPT Pro subscribers gain access to a higher-quality “Sora 2 Pro” model; ChatGPT Plus users do not receive additional benefits beyond the free tier.
The release was presented during a livestream on YouTube hosted by OpenAI Sora team researchers Bill Peebles, Rohan Sahai, and Thomas Dimson, who walked viewers through the model’s capabilities, the cameo feature, and the roadmap for upcoming tools.

A Step Forward in Video AI
It’s hard to recall now, but OpenAI wowed the world with its realistic AI video when it first teased its original Sora video model in early 2024, only to stagger the roll out slowly to a small number of creative partners until it finally released it to the public in December 2024.
By then, the entire AI video generation space had largely moved on, and has continued to advance in the subsequent months, with numerous other photorealistic AI video models from startups Runway, Luma, Kling, Higgsfield and many other competitors emerging, many with better quality and baked-in audio generation, which the original OpenAI Sora model and editor lacked — until today.
OpenAI describes the original Sora as its “GPT-1 moment” for video, a point when video generation first started to show signs of plausibility.
Sora 2, by contrast, is framed as something closer to a “GPT-3.5 moment,” marked by more advanced physics, realism, and controllability.
The model can handle complex actions such as gymnastic routines or paddleboard tricks while obeying physical rules like momentum and buoyancy.
Unlike earlier systems that might “teleport” a basketball into a hoop, Sora 2 renders a realistic rebound when a shot is missed. It also synchronizes dialogue, background audio, and sound effects, producing cohesive video-audio experiences across styles ranging from photorealistic to anime.
A standout feature is “Cameos”, which let users insert themselves or friends into generated scenes after a short one-time recording to capture likeness and voice.
Presenters emphasized during the livestream that cameo use is fully opt-in, protected by verification challenges to prevent impersonation, and revocable at any time. Here’s an example of a cameo Sora 2 video featuring an OpenAI researcher interacting with bigfoot:
The Sora App
The new Sora app serves as the primary entry point for the model. It enables users to create and remix videos, browse a personalized feed, and collaborate socially. Users can drop themselves into others’ videos through cameos, remix trending creations with their own twist, and guide style and tone through prompts.
The app is invite-based at launch, with OpenAI aiming to ensure people join alongside friends. According to the company, the feed is designed differently from typical social media platforms.
Instead of maximizing time spent scrolling, the app prioritizes discovery of videos likely to inspire creation. Content is weighted toward people a user follows or interacts with, and personalization can be adjusted through natural language instructions.
Sora is available on iOS for free, with generous usage limits subject to compute capacity. Over time, OpenAI plans to offer optional paid tiers for generating extra videos when demand is high.
ChatGPT Pro subscribers will also gain access to a higher-quality “Sora 2 Pro” model via sora.com and, eventually, in the app. An Android version is in development.
Identity Protection and Cameos
The cameo system is also central to identity protection on the platform.

Setup and verification: To create a cameo, users record a short video and dynamic audio sample in the app. OpenAI’s systems validate the sample with audio challenges to ensure authenticity and prevent impersonation.

Control over permissions: Once verified, a user chooses who can use their cameo in generations: only themselves, selected contacts, mutuals, or everyone. These permissions can be edited at any time in cameo settings.

Customization: Users can adjust how the model portrays them, correcting quirks like clothing or accent hallucinations, or adding playful stylistic variations.

Revocation and deletion rights: At any moment, a cameo owner can revoke access. They also have the right to delete any video featuring their likeness — including drafts created by others. OpenAI describes this as giving users ownership-style control over their identity within the system.

Safety for Teens and the General Audience
OpenAI has placed heavy emphasis on teen safety and wellbeing features at launch.

Anti-doomscrolling design: For users under 18, the app disables infinite scroll by default. Instead, the feed naturally pauses after a set number of videos, requiring a cooldown before resuming. Even adult users are nudged if they appear to be caught in extended passive scrolling, as the app prioritizes creativity over consumption.

Content safeguards for minors: When classifiers detect a potential minor in an uploaded cameo recording or image, stricter thresholds apply. This ensures that subsequent generations are filtered against additional categories of harmful or inappropriate output.

Privacy defaults: Teen accounts come with stronger privacy settings, limiting how their likeness can be used, restricting discovery by adults, and adding barriers to unsolicited contact.

Parental controls: Parents can use ChatGPT-linked tools to adjust a teen’s experience. These controls can override feed limits, disable algorithmic personalization, manage cameo permissions, and restrict direct messaging.

These measures reflect OpenAI’s intent to balance experimentation with wellbeing, recognizing concerns about addictive behavior and harmful social dynamics.
Safety and Provenance
Alongside these identity protections, OpenAI’s system card outlines broader safeguards:

Input and output moderation using multimodal classifiers.

Restrictions on generating public figures or photorealistic likenesses without consent.

Automated detection of harmful content, with extra scrutiny applied to Sora’s social feed.

Provenance features such as C2PA metadata, visible moving watermarks on downloaded videos, and internal tracing to verify AI-generated content.

The company partnered with external red-team testers to stress-test the system against categories like extremism, nudity, self-harm, and political manipulation.
Roadmap: Storyboards and API
Beyond the app, OpenAI highlighted new features in development for sora.com, including storyboard tools that will let creators control how a video unfolds shot by shot. According to the livestream, storyboard capabilities are expected within weeks.
The company also confirmed that an API for Sora 2 will roll out “in the coming weeks,” opening the model to developers who want to integrate video generation into their own tools and editors.
“There’s a long tail of use cases where people do amazing things, where we might not want to build fine-grained editing controls, but others might,” one presenter explained.
Altman’s Reflections
OpenAI CEO Sam Altman shared his perspective in a personal blog post accompanying the launch. He described Sora as “the ChatGPT for creativity moment” and said early testers found the cameo feature a surprisingly compelling way to connect with others.
Altman also acknowledged potential downsides. “We are aware of how addictive a service like this could become, and we can imagine many ways it could be used for bullying,” he wrote. He cautioned against the risk of an “RL-optimized slop feed,” but emphasized that the team has worked to design around that outcome.
He outlined several guiding principles: optimizing for long-term user satisfaction, giving users control over their feeds, prioritizing creation, and helping people achieve their goals. If users do not feel their lives are improved after several months, Altman said OpenAI would make significant changes, or discontinue the service entirely.
What Comes Next
For now, Sora 2 is available through the app and sora.com, with API access coming soon. Sora 1 Turbo remains active, and users’ past creations will stay accessible in their libraries.
OpenAI frames Sora 2 not only as a tool for entertainment and creativity but also as a step toward broader ambitions in world simulation and AI systems that can interact with physical reality. As the Sora team put it in their announcement, the system is still imperfect, but it signals progress toward “simulating reality” while offering users a playful new way to experiment with video.
The combination of new tools, a social app designed to prioritize creation, a robust identity protection framework, and an API aimed at developers suggests OpenAI sees Sora 2 as both a consumer product and a platform. Whether it becomes a staple of digital creativity will depend not only on its technical abilities but also on how effectively the company manages safety, wellbeing, and user trust in the months ahead.