OpenAI says Codex with GPT-6.2 Astra rebuilt a full video set in Blender from two frames. The demo starts in front of a green screen and ends with a walkable room. It claims to match camera blur and studio lighting gear from those frames alone.
The obvious reading is that set building is now a prompt. I do not buy pixel-perfect yet.
OpenAI says Astra iterated until the rebuild was pixel-perfect identical, then let the user change wall art, widen the room and relight the scene. That would have to hold on messy sets with reflections, thin geometry and moving cameras, not just a controlled green screen shoot. No render time, no code and no failure case are shown, so there is no way to check the work.
To be fair, the direction is real and useful. A rebuilt set that becomes a game-like space you can walk through, plus a webcam app that swaps static backgrounds for a live room, is a strong proof of 3D understanding. OpenAI adds that the Decisions API can match webcam lighting to the virtual room, and its guide says that API evaluates text, images and JSON and returns typed answers about 10x faster than the Responses API.
If you are building with this, budget for cleanup and test that lighting match on real calls before you promise it to clients.




