OpenAI drops Sora 2 with synchronized audio and a new social app

Remember when the first Sora dropped and video generation felt magical, yet objects would spontaneously melt whenever the laws of physics got in the way? OpenAI is ready for round two.

OpenAI officially released Sora 2, calling it the "GPT-3.5 moment" for video. Along with the flagship video and audio model, the company rolled out a brand-new, invite-only Sora iOS app designed around community co-creation.

Physics that actually work (and synced audio, too)

Older video models were notorious for being overly optimistic cheaters. If a basketball player missed a shot, the model would often glitch the ball straight through the hoop just to satisfy the prompt.

Sora 2 tackles that head-on. If a player misses now, the ball realistically bounces off the backboard. OpenAI notes that the model is much better at respecting the laws of physics, successfully handling brutal physical dynamics like backflips on a paddleboard with accurate buoyancy, or high-flying gymnastics routines. Even its mistakes look like errors made by an agent inside the simulation rather than a breakdown of reality.

Crucially, Sora 2 is no longer a silent generator. It natively produces synchronized, realistic audio, including ambient soundscapes, sound effects, and spoken dialogue tailored to the scene.

Controllability has also taken a big leap. The model can maintain consistent world state across multiple camera angles, seamlessly jumping between photorealistic, cinematic, and anime aesthetics.

Putting yourself into the scene

The marquee feature powering the new release is called "Characters." After a quick, one-time audio-video recording to confirm your identity and capture your likeness, Sora 2 can drop you, your friends, or your pets directly into any generated world with remarkable fidelity.

OpenAI built an entire iOS social network around this exact mechanic. Inside the app, users can browse a feed, generate clips, and remix each other's creations. The company is leaning heavily into social connection, hoping that remixing your friends into wild scenes becomes the next natural evolution after text, emojis, and voice memos.

Aware of algorithmic doomscrolling fatigue, OpenAI claims the app's recommender system is powered by natural language prompts and explicitly optimized for creation rather than raw watch time. Teens get default viewing limits, and parents can manage settings directly through ChatGPT.

For now, the Sora app is rolling out on iOS via invite in the U.S. and Canada, with web access at sora.com following shortly after. It launches with generous free limits, while ChatGPT Pro subscribers get access to an experimental Sora 2 Pro version.

Why it matters

Turning a frontier world simulator into a consumer social app is a fascinating, audacious play.

For everyday creators, Sora 2 essentially collapses the entire short-form production stack. Generating believable visuals, coherent camera movement, and synchronized audio in a single pass wipes out hours of post-production friction. But real-world compute bottlenecks are still real, and until waitlists clear and generation queues shrink, it remains a showcase rather than a daily utility for most.

For the AI industry, OpenAI is shifting the goalposts. While competitors like Runway and Kling have raced to woo Hollywood editors and professional visual effects studios, OpenAI is aiming straight at TikTok and Snap. Turning generative video into a collaborative, meme-friendly social playground completely alters how models are monetized and distributed.

In our view, the biggest question mark isn't the physics engine—it's safety and cultural buy-in. OpenAI gives users fine-grained controls to revoke their likeness and review any drafts featuring their Character. Still, putting photorealistic identity cloning into a viral social loop will test safety guardrails like nothing before. If OpenAI pulls it off, video generation will finally stop being a tech demo and become a shared digital language.