This might actually be quite nice - the Blender Python API is currently very useful and very touchy. Lots of differences in behavior in headless mode which are hard to debug (because you can't open the GUI to see what's happening, because that changes the behavior).
Yes the blender API feels like it sits on top of the GUI rather than the GUI on top of the API. When you are writing scripts in the blender api you basically mechanically describe the steps you would take in the UI. It can be a little fragile at times.
I've used Claude to write some blender scripts and it's an excellent use case. I look forward to even better claude/blender interaction based on this annonuncement.
I've also used genAI to write script. It works splendid up to a point, then there is absolutely no way to move the needle further. And it's not even close to renders I would ever publish.
That being said, it's about the same for the code it produces for non purely creative things, but for artistic work, I doubt an LLM in between gives any gain. After all, we do have an interface. A human interface.
Artists mad about AI art ought to welcome this. This is about making art tools better, instead of replacing them entirely. The alternative to this is AI just generating art directly and making tools like Blender obsolete.
Art generators need to come a long way to completely replace art tools. I dabble, but if I were doing real work with it, there have been times it would have been faster to composite in a 3D model rather than keep trying to prompt an image generator into fixing something.
This is what I do. It’s been really helpful for taking existing FBX files and handing them off to the agent + Python Blender API to analyze the geometry, convert to GLBs, etc.
What model you using? With codex and gpt-5.4 set to xhigh (and now gpt-5.5) seems to have zero issues helping me with rigging and fixing glb/fbx models, works as a charm. One time I instructed it to iterate together with screenshots because it was a gnarly task, but usually it figures out everything even when headless.
I disagree that anyone should need LLMs for Blender, for example, because Blender is designed by people to be understood and used by people, even if it requires a learning curve. It sounds a bit dangerous to build new things we don't understand, or worse, reduce our understanding of what we currently use because (only after studying our use of the same technology) an LLM apears able to replicate it, mostly.
I'm reminded of Sam Altman's performative helplessness on Jimmy Kimmel, when he described being unable believe a baby without ChatGPT. That's something I believe humanity has been capable of doing for a good portion of its existence, and not something we should give up to the hands of a yet-unproven, yet-unprofitable technology.
Surely there's a middle ground where improved APIs can be leveraged by both people and LLMs alike while keeping those APIs approachable? Why is it necessary that changing the python APIs would lead to "need[ing] LLMs for Blender"? I'm nowhere close to an AI maximalist but this criticism seems grounded in execution concerns. I'm definitely not saying that they won't mess this up and make the APIs overly complex, I just don't think that's necessarily going to be the case.
Regarding whether AI can/could overcome the hurdle of human understanding: I'm not sure if that's really a hurdle. Let's say in theory, a system was crafted by AI to be interacted with exclusively by AI. Broadly, I assume the outcome of the system would be for people, and it would have some purpose or value. Now my question is: how do we verify it functions? If it is a black box that nobody understands, then we can't verify it at all, and we can't debug it if there's something wrong with it. We circle back to the human understanding issue.
(I'm sorry if my tangent about Altman was taken as a personal affront, as I did not mean it to be that. It just muddied the two interesting topics you brought up.)
Not everything is abstract art. Sometimes I want my subsurf modifier to only target certain vertex groups, and if I can use AI to make that happen in a few seconds, that's a huge win for me.
Blender (and CAD programs as well) get in the way of creativity.
I know what I want, no idea how to tool my way there.
I spend two months going through YT tutorials, mucking about in Blender in order to figure out how to put together the model I have in my head [1].
(A year later, a new project idea—and it's back to YouTube because the learning is not only a steep curve but also sometimes so esoteric that it's fleeting.)
Absolutely agree - I was not impressed, but it will be a lot easier to work with the tooling without a 10 month crash-course on UI and 3D terminology if I can ask for what I want in plain language instead of knowing which button buried three levels deep to press to get my desired results.
Frankly, I love the idea of an automation engine printing out tangible works. I actually build spritesheets that way! Load a bunch of individual gimp files as layers, set them offset by a given parameter, and boom, done!
Would be rad to incorporate some statistical procedurally generated designs based on my own aparatus.
What I do not want to see is this realm of LLMs hijacking decades of hard work and consideration for integration channels to more tailor towards their LLMs, not for the diligent engineer.
If they want to put their tentacles as far as they want while making products more difficult to work with innovation of a different color, they are making enemies out of, at least me.
Honestly, I think this is a stepping stone towards replacing industry CAD modeling tools.
AI _can_ work with 3D models already, but it's really bad at it. CAD requires an extra level of control and I think this is where I could see AI companies wanting to get a foot in the door.
e.g "Let's build an adapter between 2in BSP Male and 3/4in NPT Female threads with a third Hose Barb outlet with the following properties..."
So they want claude to be able to talk to blender