A client marks up a conference room to turn it into two private offices. In a walkthrough, they want to see whether both rooms get a window. On the construction set, the questions get more precise. Where does the partition meet the window mullion? Which way does the new door swing? Will both offices appear in the room schedule?
GPT-6.2 Astra and GPT-6.2 Sol are now in ArchiLabs, including ArchiLabs Desktop. Two results from the November 17 launch are especially relevant to that revision. Astra earned the highest rating for interactive 3D scenes on AA-3D. Both models also improved substantially over their predecessors on BenchCAD, where a part can look right and still lose points because a hole stops short. OpenAI’s launch announcement
What GPT-6.2 Astra’s 1,318 Elo on AA-3D tells a designer
One of the 55 briefs in Artificial Analysis’s AA-3D benchmark asks for an interactive Colosseum. Reviewers can open two versions built from the same brief, explore them in a browser, and vote for the one they prefer. The model names stay hidden. Those votes determine the Elo rating. The higher the rating, the more likely reviewers are to prefer that model’s scene in a head-to-head comparison. On the launch-day board, GPT-6.2 Astra placed first at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259. AA-3D leaderboard and example scenes
For a design team, that result speaks to a particular part of the job. A client walking through the proposed offices can judge how the space feels from the doorway. To establish that the new wall lands on the mullion, someone still has to measure the geometry. BenchCAD gets closer to that second question.
GPT-6.2 Astra vs GPT-6.2 Sol on BenchCAD 2.1 Core
BenchCAD’s Core set contains 100 held-out hardware engineering cases. Some ask the model to build a part from a dimensioned drawing. Others supply the parts and ask the model to assemble them. The benchmark authors show a forked lever with two parallel sides. In one of the authors’ earlier examples, the submitted lever has a cross-hole in only one side. The part looks plausible until you follow the path the pin has to take. The scorer gives partial credit according to how closely the submitted geometry matches the reference. BenchCAD’s illustrated task examples
OpenAI ran the models in Codex at maximum reasoning effort. BenchCAD reports a score out of one and the cost of each benchmark case. BenchCAD 2.1 results and test conditions
- GPT-6.2 Astra scored 0.825, up from 0.626 for GPT-6 Astra. Its run averaged $5.43 per case.
- GPT-6.2 Sol scored 0.731, up from 0.526 for GPT-6.1 Sol. Its run averaged $1.60 per case.
Sol now scores above the previous Astra on the same test. That makes it an interesting starting point for everyday drafting. We’d use Sol when the markup specifies the change and Astra when several views or dimensions have to agree before anyone can place the geometry. The dollar figures describe OpenAI’s benchmark runs. ArchiLabs Desktop bills through its own subscription and usage credits.
GPT-6.2 for Revit: carry a design revision into the model
Back in the conference room, the plan might show where the partition ends while an interior elevation shows what it meets. If the two views point to different mullions, ask Astra to identify the discrepancy. A dimension from the project architect settles the question. Once the location is clear, select the room and adjoining walls in Revit, attach the marked-up plan in Desktop, and give Sol the edit.
Split the selected conference room at the dimension shown. Keep the corridor wall in place and use the door type from the office next door. Set both room names to Office and assign room numbers 214A and 214B.
Open the room schedule after checking the plan. A Revit room tag and a room schedule both read the name and number from the room itself. If two tags are visible but only one office appears in the schedule, check that both rooms are placed and enclosed, then check the schedule filters. Editing the sheet text alone wouldn’t fix the missing room. ArchiLabs supports native model editing in Revit 2023 through 2027.
GPT-6.2 for AutoCAD: turn the same markup into DWG edits
In a 2D DWG, the partition may be a pair of lines, the door a block, and each room label a separate text object. Moving the lines carries no instruction to rename the room. You can help Sol by pointing to an existing office drawn the way you want the new one to plot.
Split the conference room at the marked dimension. Use the partition layer and door block from the adjacent office. Match the door swing in the markup and replace the conference-room label with Office 214A and Office 214B.
In AutoCAD, check the block’s insertion point at the hinge and open the plot preview. A door can sit in the right opening and swing into the wrong room. A wall can look fine in model space and disappear on paper if its layer doesn’t plot. Those two checks tell you whether the edit will read correctly on the issued sheet.
GPT-6.2 for CAD beyond the floor plan
Enactra’s BuildingBench gives a third view of spatial reasoning. It supplies four images of a building and the camera positions, withholds the building height and footprint, and asks for a textured 3D model. Across the same 12 buildings at maximum effort, Astra scored 0.895 overall and Sol 0.876. Their geometry scores both rounded to 0.856. Astra’s larger advantage came in surface agreement and appearance. That breakdown matters if you’re choosing a model to reconstruct an exterior from site imagery. BuildingBench results and explorable reconstructions
For a dimensioned CAD detail, the lever from BenchCAD is the closer comparison. Supply a section through the hole as well as the face view. The face tells you where to put the hole. The section tells you how far it goes. The same habit pays off on the office revision when a floor plan alone leaves the wall-to-window connection ambiguous.
Use GPT-6.2 Astra and Sol in ArchiLabs Desktop
Bring the files that resolve the revision into ArchiLabs Desktop on Windows. You can attach the marked-up PDF, select the walls in Revit or the objects in AutoCAD, and include the room schedule from a connected project account. Once the team settles the mullion location, ask Desktop to make the native edit. The next person opening the file can move the wall, inspect the door, or issue the drawing. See ArchiLabs Desktop at work
Desktop starts at $19 per month plus usage credits. See pricing or try Desktop with your next markup.
