Sora 2 shipped this week with native audio generation and, more importantly for how I actually use it, physics that finally look like physics rather than a rough guess at physics. I spent the first few days throwing real client-adjacent prompts at it instead of the cherry-picked demo prompts everyone posts on launch day, and the gap between the demo reel and daily usability is smaller than usual.
Audio is the actual headline
Every previous video model I've used required a separate pass for sound design, foley, even basic ambient noise. Sora 2 generating synced audio directly, footsteps that land when a foot lands, ambient room tone that matches the visible space, means I'm cutting a step out of my pipeline for certain shot types. It's not replacing a real sound designer on anything client-facing yet. For quick concept reels and social cuts, it's a genuine time save.
Where physics improved and where it didn't
Fabric movement, hair, and liquid all look noticeably more grounded than the previous generation, less of that telltale floating quality. Cloth actually seems to respond to the body underneath it moving rather than draping independently, which was one of the more obvious tells in earlier models. Fast, complex multi-object interaction still breaks down, though, a person catching a thrown object still tends to look approximately right rather than physically correct if you watch closely.
- Real win: synced audio cuts a full post-production step for quick content
- Real win: cloth and hair physics hold up under closer scrutiny now
- Still weak: fast object interaction between multiple moving elements
What this means for a working pipeline
I'm not switching everything to Sora 2 overnight. What changes immediately is which shots I attempt with AI generation versus which I still book a real actor and location for. Anything requiring precise interaction between a person and a physical object stays on the real-shoot list. Ambient scene-setting shots, establishing footage, mood pieces, those move toward AI generation faster now that the audio layer means less manual work afterward.
A week isn't enough to fully judge a model, and I expect rough edges to surface as more people push harder prompts through it. First impression stands, though: this is the first release in a while where the audio feature alone justifies switching tools for a specific chunk of my workflow.