GPT-6 Astra's 3D modeling capabilities are way beyond what I expected

I've been seeing several demos of GPT-6 Astra being used for 3D modeling, and there's something interesting about the way it approaches the task. It doesn't just seem to generate a model from a prompt. In many of the examples I've seen, it can reason about the geometry, make targeted modifications, inspect the result, and iterate on it. What I find particularly interesting is the iterative part. Previous AI 3D workflows often felt like: generate → fix manually → generate again. Astra seems much closer to an actual workflow where the model can identify problems and progressively improve the result. I'm wondering what changed under the hood to make this possible. Is this mainly better vision + reasoning, better tool use, or something more specific to how Astra was trained? For those who've experimented with it, what do you think is actually driving this jump in 3D modeling?