Automating 3D Modeling with GPT 6 Astra in Rhino
This mini course with Dusan Cvetkovic puts the new GPT 6 Astra model to work in Rhino. We prompt Codex to research ten famous buildings, then let it write the Python that builds each one step by step in an open Rhino session, roof last. You see the prompts, the build times, the mistakes and the grade every project earns, from a 23-minute Farnsworth House to a parametric Serpentine Pavilion built directly in Grasshopper.
In this mini course, we are going to put the newest GPT model, GPT 6 Astra, to work on real architectural modeling in Rhino. Instead of modeling by hand, we prompt Codex, let it research each building, and then watch it write Python through RhinoScript and build the geometry element by element inside an open Rhino session, with the roof added last so the whole process stays visible.
We go through ten famous buildings, from easy to hard: Farnsworth House, Church of the Light, LEGO House, Villa Savoye, the National Assembly in Dhaka, the Niteroi Contemporary Art Museum, the Serpentine Pavilion 2016, Olympic House, Shanghai Tower and the Heydar Aliyev Center. For every one of them you will see the research report Astra puts together, the exact prompts, how long the build took, what it got right, where it failed, and the grade it earned.
Along the way you will learn how to set up the Codex app and the Extra High reasoning mode, how to give the model a project folder with reference images, how to prompt for a visible step by step build, how to add a quality control pass for gaps and intersections, and how to feed it comparison screenshots so it corrects its own mistakes.
The Serpentine Pavilion is built as an editable Grasshopper definition instead of a static model, and Shanghai Tower is modeled as structure and facade only, with blocks and checkpoint files to keep a 470 MB tower under control. The Heydar Aliyev Center is the one honest fail, and the whole diagnosis and second attempt stay in the recording, because that is exactly where the limit is right now.
This mini course is best suited for anyone who wants to see what AI can and cannot do in Rhino today, and how to use it as a starting point for conceptual models instead of modeling everything from scratch. No prior AI experience is needed; basic Rhino knowledge helps.
- How GPT 6 Astra actually drives Rhino: Python written through RhinoScript and executed live in your open session, with no mouse automation
- How to set up the Codex app, pick the Astra model, choose Extra High reasoning, and manage usage limits on the ChatGPT plan
- How to give the model a project folder with reference images and prompt it to research plans, sections, photos and CAD before it models anything
- How to prompt for a visible build: element by element, floor by floor, interiors before floor plates, roof last, so you can record and check the process
- How to add a quality control step so the model checks its own gaps, intersections and unfinished facades before calling a project done
- How to correct the model with comparison screenshots and orthogonal views, and why a direct diagnosis prompt works better than a vague one
- How to ask for blocks, Boolean difference and no furniture to keep a 400 MB tower file manageable
- Where Astra stands today: what it nails (orthogonal, stacked and rule-based geometry, a full Grasshopper definition) and where it still fails (Zaha-style freeform shells)
Dusan Cvetkovic
Founder of How to Rhino and the lead behind RhinoFrame. Dusan has spent more than a decade teaching Rhino and Grasshopper to architects and building tools that bring computational design and AI into everyday architectural workflows, reaching a community of 89,700+ YouTube subscribers.
