You are viewing a flat. You go into the living room, through to the kitchen, take a look at the bedroom, come back out through the hallway.
Ten minutes later, out in the street, you could sketch the floor plan on the corner of a napkin. Not perfect, but not far off: the kitchen opens onto the living room, the bedroom is at the back, the bathroom right next to the entrance.
You have never asked yourself how you do this. For a machine, it is a real headache. And a new test has just measured where they stand.
The exercise
The test is called Blueprint-Bench. The principle is simple: an AI is given a series of photos taken inside a flat, room by room. It has to work out the complete floor plan, seen from above.
In the second version of this test, it is Claude Opus 5.5, the model released on Tuesday by Anthropic, that comes top of the ranking.
Coming first in a test does not mean you will be able to use it tomorrow to furnish your living room. But it says something interesting about what these systems are starting to understand.
Why it is so difficult
Think about what is actually required to go from photos to a floor plan.
Linking rooms you never see together. In each photo, you only see one room. You have to work out that the door to the left of the kitchen leads to the hallway glimpsed in another image.
Guessing what is not photographed. The thickness of the walls, a cupboard, a nook behind a door. The plan has to hold up even with gaps in the information.
Recovering the real dimensions. A photo flattens depth. A wide-angle lens turns an ordinary hallway into an endless tunnel. The proportions have to be reconstructed from clues: a door, a bed, the size of a floor tile.
You do all of this effortlessly, because your brain has spent your whole life moving around buildings. An AI that has mostly learned from text and images has never walked anywhere.
This test is not really about drawing floor plans. It measures something deeper: does an AI understand that the world has three dimensions, with a top, a back, rooms that touch each other? This is precisely what robots lack in order to move around your home. On Saturday we explained why cleaning robots still do not exist: they do not understand the messy space of a house. A model capable of reconstructing a flat from photos is one piece of that puzzle.
What it could be used for
The uses are easy to imagine. A property listing that automatically generates a floor plan from the seller's photos. A tradesperson who estimates a renovation without travelling. A house move prepared by knowing in advance where each piece of furniture will go.
And further ahead: robots capable of finding their way around a home they have never seen, simply by looking around them.
What to keep in mind
An approximate floor plan is perfectly fine for a listing. Much less so for ordering a made-to-measure kitchen to the nearest centimetre. No one claims to be there yet.
But look at how far things have come. Not long ago, these systems could describe a photo: "a living room with a grey sofa". Today, they are starting to reconstruct the space that exists around it.
This is the kind of progress that makes no noise, and that matters more than many spectacular announcements. Opus 5.5 has mostly been talked about for its price. Perhaps its most interesting advance is hidden here.