Note
November 19, 2026
2 min read
My Astra 6.2 Eval Had a Creek
I gave GPT-6.2 Astra one prompt for a walkable Amazon rainforest. That first build gave me a way to start running my own simulation experiments.

GPT-6.2 Astra just came out, and I’ve already seen plenty of eval scores for it. They tell me how it did on different tests. I wanted to see how it would work the way I work, starting with a rough idea and refining it once I had something I could try. I gave Astra one prompt asking it to build a 3D Amazon rainforest I could walk through.
The first version of Igarapé had a creek, vegetation and wildlife rendered in real time in the browser with JavaScript and Three.js. In later passes, we added terrain that generates as you explore, weather that changes, and animals that respond to what’s happening around them.
I could walk through a version, point to something that needed work, and have Astra research it, revise the code and run it again. For the toucan’s fruit toss, Astra used measurements from high-speed studies of related species to work out how quickly the bill should open and how the fruit should travel. The build, research and revisions used about 15% of my Pro 500 usage allowance.
I’ve wanted to do more work with simulations for a while. I can now explore the forest in the browser or run the same simulation unattended in Node.js. I want to run the same day twice, bring the rain in an hour earlier on the second run, and see what the animals do differently by dusk.
This writing reflects my personal perspectives on product management, AI, and content discovery. It does not represent the official position of my employer or any affiliated organization.
