OpenAI quietly shipped something interesting in GPT-6 Astra: it can generate actual 3D models, not just talk about them.

Simon Willison has been testing it with Blender on macOS, and the workflow is absurdly simple. Install Blender from blender.org, then prompt ChatGPT with something like “Use the already installed /Applications/Blender to render a scene of a pelican riding a bicycle.”

It works. The model writes Python scripts that drive Blender’s API, sets up scenes with geometry and materials, positions cameras and lights, and spits out rendered images. Willison followed up with “add a background and a lot of flair” and then “make it a whole lot better,” and got progressively more elaborate pelican-on-bicycle scenes.

This isn’t a toy feature. According to OpenAI’s developer announcement, Astra “excels at building 3D models” and developers have used it to create “incredible renderings of gardens, shipyards, animals, cityscapes, even Dyson spheres.” The model has better prompt understanding and attention to detail than previous versions, which matters when you’re generating procedural geometry where a misplaced vertex ruins everything.

What This Actually Means

The interesting part isn’t that an LLM can call Blender’s Python API. You could get GPT-4 to do that if you fed it enough documentation. The interesting part is that Astra apparently understands 3D space well enough to generate coherent scenes without you specifying every detail.

When Willison asks for “a lot of flair,” the model has to make aesthetic decisions about composition, materials, and lighting. When he asks to “make it a whole lot better,” it needs to evaluate what “better” means for a 3D scene. That requires some kind of internal model of how 3D graphics work, not just pattern matching on code snippets.

This matters for robotics too. There’s a separate OpenAI page specifically about GPT-6 Astra on robot arms, suggesting the same spatial reasoning that helps with 3D modeling might transfer to motion planning and manipulation tasks.

The Developer Experience

The macOS setup is particularly clean because Blender installs as a standard app bundle with a Python interpreter and all the libraries you need. The coding agent just points at /Applications/Blender and starts writing scripts. No Docker containers, no virtual environments, no dependency hell.

This is probably not how most people will use Astra’s 3D capabilities in production. But it’s a good signal for how accessible this stuff is becoming. You can go from “I wonder if this works” to rendered 3D output in a single prompt.

The pelican thing is also telling. Willison’s been using that as a test case across different models and tools, and Astra apparently handles it well enough that OpenAI’s own promotional video at the 1:59 mark shows a pelican with a red neckerchief. Either they’re reading his blog or the model really did independently decide that’s a reasonable interpretation of the prompt.

The Bigger Pattern

We’ve seen LLMs get better at code, better at math, better at following multi-step instructions. Now they’re getting better at spatial reasoning. That opens up a whole category of tasks that were awkward to automate before.

Game asset generation, architectural visualization, procedural level design, robot motion planning, anything where you need to reason about objects in 3D space. These were all domains where you needed either specialized tools with steep learning curves or expensive contractors. Now you can just describe what you want.

The code quality still matters. An LLM that writes bad Blender Python is worse than useless because debugging someone else’s procedurally generated spaghetti is miserable. But if Astra can consistently generate working, readable scripts, that changes what’s practical to build.

And if it can’t, well, at least we’ll get some entertaining pelicans out of it.