Are models as bad as humans in the spatial orientation test?

In short

They are really good and within a few degrees of the answer.

A café table seen from above with seven labelled dots: menu, flower vase, sugar bowl, water glass, espresso cup, croissant and milk jug.
The table every run was given. Each question picks a seat, a thing to face and a thing to point to.

Finding

The Spatial Orientation Test asks you to imagine standing at one object, facing another, and point to a third. People get worse the further they have to turn from the way the page faces them. I made a café version with seven things on a table and ten questions, two at each turn from 0 to 180 degrees, and told the models to answer by imagining it.

Opus 5.5 and DeepSeek 4.1 Flash were within a few degrees in all the runs.

Opus and DeepSeek did the trigonometry, and DeepSeek used code in 3 runs when told not to.

The 30 runs cost $0.58.

By email

Older, 11 Oct

Does a model know a knit stitch is wider than it is tall?

Newer, 14 Oct

Does Gemini 4 Argon know the coffee family tree?