Pat Simmons uses GPT-6 Astra with Blender to attempt three projects: a detailed Fallingwater recreation, a measured model of his studio and a 3D 1X Neo humanoid robot embedded in a redesigned product page.
The recurring workflow separates research from execution. One agent gathers referencesAI agent delegation assigns a bounded task, context, authority, and expected result from one participant to an agent or another agent. and writes a detailed PRD, then a fresh build agent controls Blender through an MCP bridge and Python APIAgent tool use lets an agent call bounded external capabilities so it can retrieve information or change state beyond generating a response., iterates for hours and checks its own work. This is a creator-run demonstration, not a controlled or independently reproduced benchmark.
The results show strong persistence, research and iterative refinement, but they also expose important limits. The Fallingwater tour contains rendering artifacts, sparse rooms and imperfect water, the studio includes boxy or inaccurate objects, and the robot has small color differences. The longest builds run for roughly 20 to 29 hours, while the studio takes about three and a half hours.
Pat Simmons reports approximately $1,700 in equivalent API cost across the three projects, while noting that the estimate is derived from logged input and output tokens and is not exact. He completed the work under a ChatGPT Pro subscription and used a banked usage reset during the second build.
Watch on YouTube
