The machine draws. Who checks?

On 3 September 2026 OpenAI released GPT-6 under the name Astra. Most of the coverage was about parameters and benchmark tables. The clips that went around among engineers were about something else: the model was operating Blender. Not describing a house. Building one. It writes the script, runs it, renders the scene, looks at the render and corrects its own work. Within a week, people with early access had posted five demonstrations that anyone in our industry should watch twice.

The demonstration that matters is the least pretty one

The house from a text brief is impressive and the palace is beautiful. The one to study is the tyre. A raw scan goes in. The model asks a question back, hollow rubber or solid, then works alone in Blender for half an hour and writes a STEP file. SolidWorks opens the file without complaint. That last sentence is the whole story. A STEP file is not a picture. It is a precise, dull, machine-readable description of geometry that a different program, written by different people, accepts or rejects. The model did not persuade a viewer. It satisfied a parser.

That is the pattern we have bet the company on, seen from the other side. A language model is a very good proposer and a poor authority. Let it propose in a form that a deterministic system can accept or refuse, and you get the model’s range with the system’s rigour. Astra writing scripts instead of clicking, and STEP instead of screenshots, is that pattern at planetary scale. We build the same shape for documents and plans: the model reads and proposes, the graph stores and answers, and every fact keeps the page it came from.

What nobody in the clips asked

Watch the demonstrations again and note what the human does. In every one, the person judges the result by eye. Does it look like the house. Does the locomotive look right. Nobody asks whether the door on the second floor is a T30 fire door, whether the duct crosses the beam, or whether the 3,295 objects in the locomotive match the drawing’s third revision or its fourth. Generation is judged by appearance. Construction is judged by conformance. Those are different tests, and passing the first says nothing about the second.

We have a number for the second test. In August we put five vision models, from the cheapest to a frontier model at ten times the price, in front of 242 pairs of plan sheets: the same floor in two issues, with a known set of changes between them. Asked what had changed, the best model named seven percent of it. The worst named almost nothing. The same family of models that builds a palace overnight cannot tell you that five sanitary fixtures vanished between issue four and issue five.

What changes for the building industry

Three things, in our reading. First, 3D models get cheap. A model from a photo, a model from an archive drawing, a model from a scan: a task that took a modeller a week now takes a prompt and an evening. Second, that abundance cuts both ways. A project that today has one model of uncertain quality will soon have several, of unknown provenance, produced by whoever had a laptop. Which one is binding is not a question the generator answers. Third, the plan of record does not go away. The document a contractor signs, an inspector reads and a court later opens is still the drawing set, and it will be for years. If anything, the gap between a beautiful model and a binding drawing widens.

That is why we did not build a generator. We built the check. Our thesis for construction has been that you do not need a 3D model to find the conflict between a plan and a rule. After 3 September we would put it differently: you are about to have more models than you can trust, and what you need is the thing that reads them against the plan that counts.

The honest part

It would be foolish to pretend the ground is not moving. A model that operates Blender today will operate a BIM authoring tool tomorrow, and it will read a plan sheet better than the models we measured in August. We expect that, and we build so that a better reader drops in without changing the rest. What does not change is the shape of the problem: a proposal is not a proof. The tyre demonstration shows the model writing a file that another program can verify. Our job is to be the program that verifies.

Sources: OpenAI, “GPT-6 Astra: A new generation of intelligence”, 3 September 2026. Posts on X by Thomas Ricouard and Sharif Shameem (OpenAI), Tom Krcha and alpha_rover, 3 to 5 September 2026. GIGAZINE, “GPT-6 Astra 3D modeling examples”, 7 September 2026. The parameter count of roughly ten trillion is an industry estimate; OpenAI has not published one.