The most quoted number from Astra's launch week was 99.9% on ARC-AGI-3, which turned out to be harness-dependent. The number worth quoting instead is 3,295, and you can watch where it came from.
Everything below is the original post. Play them.
A drawing of a steam train, rebuilt as 3,295 parts
So how good really is GPT-6 Astra at 3D modeling?
— Tom Krcha (@tomkrcha) September 4, 2026
I took an old drawing of a steam train, gave it to Astra to reconstruct it in Blender. After few minutes it crafted 3,295 fully editable detailed objects with beautiful geometry. You can obviously tell it how detailed or… pic.twitter.com/ocEuG7QB2E
The word doing the work is editable. Most 3D generation returns a single fused mesh: you can look at it, you cannot select the chimney or change a wheel diameter, and a modeller receiving it starts over. An object tree is a different asset class. It rigs, it exports, and it survives a second person opening the file.
It takes control of Blender
The second shift is easy to miss. Astra is not writing you a Blender tutorial. It drives the application.
I guess we're allowed to talk about it now... Astra has absolutely blown my mind. I did get early access to test it out and it's definitely felt like the biggest leap we've seen in a long time!
— Matt Wolfe (@mreflow) September 3, 2026
I asked it to make me a humanoid wolf in Blender, it used computer use, took control… pic.twitter.com/NFW9K073EJ
Four words and a reference image
Since some people are asking aboutethe Codex Micro it was all done in Blender via code. The prompt was a simple "Make this in Blender" with the second image here as attached as reference. It was that easy. pic.twitter.com/2WhppBQxVy
— Pietro Schirano (@skirano) September 3, 2026
Photographs to a walkable house
GPT-6 Astra is a beast. Give it a Zillow listing. It can 3D model the house based on the listing photos and create a cool promotion video.
— Yunfan Ye (@realYunfanYe) September 3, 2026
The video was just created in one shot and there are still some wrong details. But I am sure it will be much better if I ask it to polish… pic.twitter.com/Jfok3EVACA
Worth noting that Yunfan Ye posts his own caveat in the same breath: "the video was just created in one shot and there are still some wrong details." That is the honest version, and it is more useful than the reels.
GPT-6 Astra modeló una casa completa en Blender y la convirtió en un recorrido en Unreal Engine 5. Sola.
— Goofy (@goofyninjaaa) September 3, 2026
Alberca, jardín, sala, cocina. Así vas a ver tu casa antes de que exista.
Eso hoy lo cobra un estudio de visualización con arquitectos y artistas 3D.
OpenAI dice que Astra… pic.twitter.com/Q316EMD1cm
Head to head against Claude Fable 5.1
Built a villa scene in Blender with Fable 5.1 and GPT-6 Astra
— Karan (@karankendre) September 3, 2026
This is what it built https://t.co/1UoqUd4Lyr pic.twitter.com/xV4qtJ6rQr
What it is still bad at
A detailed teardown from yage.ai, which ran exploded views, rigging and webcam motion capture, found character modelling "still leaves plenty of room for improvement, as generated figures tend to stray considerably from the original designs," that getting a character right "requires an exhausting amount of manual micro-management," and that scene assets came back noticeably low-poly.
Structure yes. Fidelity sometimes. Character design not yet.
The operator read
Ignore the renders. The transferable lesson is the object tree.
When you are choosing between two AI outputs, prefer the one that returns structure over the one that returns a result, even when the result looks better today. A fused mesh and an object tree can render identically. Only one of them is still useful after the person who prompted it leaves.
That test applies well outside 3D. A generated report versus the query that produced it. A generated design versus its component tree. A generated answer versus the pipeline that derived it. Same choice every time.
Keep reading.
Teardowns, benchmarks and playbooks on building AI-native products.
Browse the Playbook →Part 1 of Astra, Week One.
