GPT-6 Astra is here. Is it Blending?
AI
I haven’t blogged about AI for a while because gen AI has become kind of boring. These days it feels less like a tool to augment capabilities and more like a tool to replace them. For example, I recently used Claude Code to prototype a computer game idea I had, but Claude kept suggesting better game ideas than the one I’d thought up. Make your own damn game, Claude!
By taking Claude’s suggestions, the game concept improved, but it wasn’t mine anymore. We used to worry about LLMs being sycophantic. Now I’m the sycophant. Yes, Claude. Great idea. Do that.
GPT-6 Astra is here, and it’s better than ever at offloading our thinking so we can free up our brains for… something else, presumably. The promotional trailer features a guy sitting on his couch telling ChatGPT to make a spaceship computer game. ChatGPT makes a spaceship game.
Yay… but wait, where did the spaceship come from? Apparently, Astra can make 3D models in Blender. That seemed interesting.
As soon as I put my kids to bed, I fired up Astra to see how it was making 3D game assets. The promo page leans hard on “computer use,” and both the page and the video show an open Blender window, mid-edit. The implication is that Astra is in there modeling, like a person.
To try it out, I asked Astra to make and animate a 3D model based on a picture of my dog, Sinbad.
Sure enough, five minutes later I had a fully rigged game asset of my dog, albeit a semi-horrifying interpretation of him.
Did Astra control the mouse and keyboard to craft 3D objects in Blender (“computer use”), as the promo led me to believe?
No. Blender has a Python API, and that’s what Astra is using. It’s writing code that builds the model, not pushing vertices around with a mouse. Coding agents have been doing this since before Astra, for example with Opus.
Even when I tried to force Astra to utilize computer use for Blender, it opened a Python scripting window inside Blender, which, to be fair, makes sense, because it would be really inefficient to use the mouse and keyboard when a Python API is available.
What about something simpler, like drawing a picture?
Is this a breakthru or a party trick? I’m not sure. Some (who have invested trillions in AI) are declaring this improved ability of agents to pilot software, draw pictures, and play Portal AGI. While impressive, it also appears to be an inefficient and expensive way to complete tasks on a computer, and as illustrated with the Blender example, Astra itself avoids computer use when an alternative like a Python API exists. The real power probably comes from blending (no pun intended) computer use with traditional coding, which lets the agent see and test whatever it’s building.
The interesting thing about Astra isn’t that it can push a mouse around. It’s that it can run the thing it built, play with it, and decide whether it’s any good. That was also the last part I was still doing in my agentic workflows. My days of being useful to my agent feel numbered. Yes, Astra. Great idea. Do that.