Playing around with Opus 5.5 + Blender has yielded some crazy results, but my $200 plan maxes almost instantly. Any open models that are somewhat comparable (maybe like 80% as good) right now?
Any of the big ones should be fine. What I would recommend is finding non-imageful ways to accomplish tasks. This will be more reliable and consume far fewer tokens. (I'm assuming by computer use, you mean the kind that takes screenshots and drives the mouse/keyboard)
Blender has a python integration, though I have only toyed with it once or twice myself. This would be what I look at first if/when I get into Blender again. I imagine one still needs screenshots/renders with Blender, but not to drive the computer as much as close the feedback loop for implementation-outcome.