SpatialBench

The best AI models still can't read a Rubik's cube.

SpatialBench measures how well AI models understand space: reading a Rubik's cube, following a paper fold, taking a measurement off a drawing. A 3D engine generates every question, so the correct answer is known exactly and can't be memorized from the web.

On ordinary benchmarks the top models sit within 2 points of each other. On these, they range from 14% to 85%.

See the leaderboard Run it on your model
Every scene is generated, so the exact answer comes with it
rendered Rubik's cube
CubeBench · live State tracking
Sticker at row 2, col 3 of the top face? → green
folded paper, punched, then unfolded to show holes
FoldBench · in progress Folding
Fold, punch a hole, unfold. How many holes? → 4
136 76
PrecisionBench · in progress Metric precision
Center-to-center hole pitch? → 76

Rendered, not scraped

Scenes come from a physically-based renderer, so the label is the scene graph itself. No annotation noise, no web contamination.

Verified, not judged

Every answer is checked by string equality against renderer state. Accuracy and format-failure rate are published side by side.

A seed, not a file

Packs are regenerable from a seed at any size or difficulty. Contaminated? Bump the seed and re-render.

The benchmarks

One skill each

Each benchmark below tests a different piece of spatial understanding: tracking an object as it turns, following a shape as it folds, reading exact measurements off a drawing. They share one format, so scores are comparable and every answer is checked automatically, never by a human.

Live · 1,040 items · 19 models

CubeBench

Can a model read a Rubik's cube from a photo? Which color sits on a given square, what's on the face it can't see, and what the cube looks like after a turn. Rendered cubes, exact answers.

Full leaderboard →
rendered Rubik's cube
In progress

FoldBench

Fold a square of paper step by step, then ask what the model can't directly see: which layer ends up on top, where a punched hole lands when you unfold, whether it lies flat. Answers come from a physics-checked fold simulator.

In progress

PrecisionBench

Read engineering drawings the way a machinist or architect does: hole spacing, overall size, which 3D shape matches three flat views, how many fixtures on a floor plan. Every drawing is generated, so the exact measurement is known.

Planned

Permanence

Track an object after it goes out of sight. Cups and a ball, boxes that swap places: where did it end up? Easy for people, hard for models.

Planned

Perspective

Show a scene, then ask what it looks like from another angle, or whether two views are of the same place. Pure viewpoint reasoning.

Planned

The real-time track

A version of each benchmark where the clock keeps running. The scene moves whether or not the model has answered, and late answers don't count. Built for models that watch a live video stream.