OpenAI launches Sora 2, its 'GPT-3.5 moment' for video

OpenAI is betting that the path to AI that understands the physical world runs through video, and today it’s betting big. The company just released Sora 2, a model it’s calling the ‘GPT-3.5 moment’ for video generation, while simultaneously pulling the plug on the original Sora product. It’s a clean break from a model that, while groundbreaking in February 2024, now feels like an early proof-of-concept.

Sam Altman’s team is making a specific and ambitious claim: Sora 2 can handle physics in ways prior models simply couldn’t. They’re drawing a sharp contrast with what they call ‘overoptimistic’ systems that cheat to satisfy a prompt. The classic example? A missed basketball shot that magically teleports into the hoop. Sora 2, the company insists, will let that ball clang off the rim. The model’s errors are now more like the mistakes of an ‘internal agent’ failing to execute, rather than a video model warping reality. This ability to model failure, not just success, is what they’re banking on for building a useful world simulator.

The release isn’t just about raw model capability; it’s wrapped in a new consumer product. OpenAI is launching a social iOS app simply called ‘Sora,’ built entirely around Sora 2. The hook is a feature called ‘characters’ that lets you upload a short video and audio recording of yourself and then drop a remarkably accurate avatar into any generated scene. It’s a social play that the company says has already led employees to make new friends during internal testing. The app is designed for remixing and creation, with a feed that prioritizes content from people you know and videos meant to inspire your own, explicitly avoiding optimization for time spent scrolling.