Sora 2 System Card
OpenAI ● Covered by 3 sources
OpenAI dropped Sora 2, a video model that now nails physics and adds synced audio to your clips. It's the jump from 'neat AI demo' to something that could actually fool you.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI just published the system card for Sora 2, and the headline upgrade isn't just sharper pixels. It's physics. Earlier video generators were notorious for objects melting into each other or gravity taking the day off. Sora 2 is built to handle motion and collisions in a way that looks like it understands the world it's rendering, not just mimicking pixels that came before it.
The other big shift is audio. Sora's first version was mute — you got a silent clip and had to layer sound on yourself. Sora 2 generates synchronized audio alongside the video, meaning dialogue, ambient noise, and effects now arrive baked into the output rather than bolted on afterward. That's a meaningful technical jump: audio-visual sync is one of those problems that sounds simple until you actually try to make footsteps land on the right frame.
OpenAI is also pushing steerability and stylistic range, meaning the model is more responsive to specific prompts and can produce a wider variety of looks — not just photorealism, but different visual styles depending on what the creator asks for. Combined with the physics and audio improvements, this points toward a model aimed less at flashy tech demos and more at actual production use, where consistency and control matter more than a single impressive clip.
The fact that OpenAI released a full system card alongside the model — the kind of document usually reserved for safety-sensitive releases — signals they know exactly how far synthetic video has come. A model that gets physics right, syncs its own audio, and follows instructions closely is a model that produces content indistinguishable from something filmed with a camera. That's the real story buried under the spec sheet.
My take — AI-written commentary, not fact-checked reporting
I'll say the obvious thing nobody at OpenAI wants headlining their blog post: a model this good at synced audio and realistic physics is a deepfake engine with a nicer UI, and the system card reads more like a legal hedge than a safety plan. Watermarking and provenance tools need to ship alongside this, not as an afterthought six months later, or we're going to spend 2025 arguing about which viral video is real.
Read more about this at: OpenAI