OpenAI’s GPT-6 Astra can turn drawings, photographs and written instructions into 3D models, and architects are testing it with Blender, Rhino and Revit. Experiments range from detailed reconstructions to models with basic planning errors. The central challenge for architecture, engineering and construction (AEC) firms is whether these systems can create accurate, editable building information models (BIM) and keep them usable as designs change.

Astra can interpret references, plan modelling steps, write code, inspect results and make corrections. In a Revit apartment-layout test, BIM Pure’s Nicolas Catellier found that the first attempt put a fridge in a circulation path; added instructions improved the layout, but he still judged it inferior to a human-made scheme. In a separate cabin test, Astra recreated a recognisable building from photographs but used DirectShape geometry, which lacks the normal editing behaviour of Revit building elements. A plugin developer reported that Astra created 327 native walls, 162 doors and 98 windows in a 16-flat block from one screenshot, though that result has not been independently verified.

The software connection determines what an AI agent can create and edit. Through a Revit MCP server (a connection that lets software access application tools), Ideatura classified parts of its Leadenhall Building reconstruction in IFC, allowing a structural member to be identified as a column rather than only shown as a solid. In a sketch-to-detail test, Astra asked questions and produced a draft, but a reviewer still needed to correct framing, furring and component choices. OpenAI’s architectural visualisation example began with a written brief and developed through reviews into a proposed plan and furnished Blender scene; OpenAI said professional review would be needed before it could inform construction.

Keeping later revisions intact is another challenge. Augmenta’s electrical-routing engine can solve a whole building overnight, but customers previously edited results by hand after design changes. In a June 2026 interview, CPO Aaron Szymanski said version 2 adjusts affected routes and reruns only changed areas. The article says automated modelling likewise needs to build on architects’ edits instead of returning to an earlier proposal.

Early tests show both costs and measurable results. BIM Pure’s detail experiment used roughly $10 in API-equivalent tokens and took about 20 minutes for an initial recreation; the 16-flat block took 79 minutes at maximum settings. Signal’s document-checking experiment found 13 of 14 planted defects in a 40-page set and produced 23 observations, 21 confirmed. The article says these tests do not compare total effort with existing workflows, including briefing, checking, failed runs and repairs.

Large projects also test whether agents can access and coordinate information across models. Perkins&Will’s Nick Cameron said a large hospital has to be split into around 30 linked Revit models to remain workable. If an AI connector exposes only the active model, an agent may miss consequences in other project files. The article names different data approaches from That Open Company’s Fragments, Motif, Snaptrude and Qonic, while noting there is no evidence here of their performance through a large hospital project.

In the BenchCAD programmatic CAD benchmark, vendor-reported scores for mechanical-part geometry were 0.962 for Anthropic’s Opus 5.5 and 0.959 for Astra; the benchmark is not specific to AEC. The article says adoption is likely to begin with work that firms can specify and check, such as standard components governed by established rules, while detailed models developed from sketches and written descriptions remain a possible longer-term use.

OpenAI’s GPT-6 Astra can turn drawings, photographs and written instructions into 3D models, and architects are testing it with Blender, Rhino and Revit. Experiments range from detailed reconstructions to models with basic planning errors. The central challenge for architecture, engineering and construction (AEC) firms is whether these systems can create accurate, editable building information models (BIM) and keep them usable as designs change.

Astra can interpret references, plan modelling steps, write code, inspect results and make corrections. In a Revit apartment-layout test, BIM Pure’s Nicolas Catellier found that the first attempt put a fridge in a circulation path; added instructions improved the layout, but he still judged it inferior to a human-made scheme. In a separate cabin test, Astra recreated a recognisable building from photographs but used DirectShape geometry, which lacks the normal editing behaviour of Revit building elements. A plugin developer reported that Astra created 327 native walls, 162 doors and 98 windows in a 16-flat block from one screenshot, though that result has not been independently verified.

The software connection determines what an AI agent can create and edit. Through a Revit MCP server (a connection that lets software access application tools), Ideatura classified parts of its Leadenhall Building reconstruction in IFC, allowing a structural member to be identified as a column rather than only shown as a solid. In a sketch-to-detail test, Astra asked questions and produced a draft, but a reviewer still needed to correct framing, furring and component choices. OpenAI’s architectural visualisation example began with a written brief and developed through reviews into a proposed plan and furnished Blender scene; OpenAI said professional review would be needed before it could inform construction.

Keeping later revisions intact is another challenge. Augmenta’s electrical-routing engine can solve a whole building overnight, but customers previously edited results by hand after design changes. In a June 2026 interview, CPO Aaron Szymanski said version 2 adjusts affected routes and reruns only changed areas. The article says automated modelling likewise needs to build on architects’ edits instead of returning to an earlier proposal.

Early tests show both costs and measurable results. BIM Pure’s detail experiment used roughly $10 in API-equivalent tokens and took about 20 minutes for an initial recreation; the 16-flat block took 79 minutes at maximum settings. Signal’s document-checking experiment found 13 of 14 planted defects in a 40-page set and produced 23 observations, 21 confirmed. The article says these tests do not compare total effort with existing workflows, including briefing, checking, failed runs and repairs.

Large projects also test whether agents can access and coordinate information across models. Perkins&Will’s Nick Cameron said a large hospital has to be split into around 30 linked Revit models to remain workable. If an AI connector exposes only the active model, an agent may miss consequences in other project files. The article names different data approaches from That Open Company’s Fragments, Motif, Snaptrude and Qonic, while noting there is no evidence here of their performance through a large hospital project.

In the BenchCAD programmatic CAD benchmark, vendor-reported scores for mechanical-part geometry were 0.962 for Anthropic’s Opus 5.5 and 0.959 for Astra; the benchmark is not specific to AEC. The article says adoption is likely to begin with work that firms can specify and check, such as standard components governed by established rules, while detailed models developed from sketches and written descriptions remain a possible longer-term use.