The Shadow Test: frontier models build cube sculptures whose three shadows spell a word
The author designed a test where models fill a 10³ grid with cubes so three projections spell a word; GPT-6 Astra and Opus 5.5 hit the proven optimum of 92.4% on all three tries, while Fable 5.1 reached only 45.5%.
Introducing: "The Shadow Test" for Frontier AI! 🧊 I made AI models build sculptures out of cubes. Then I turned on gravity and it turned out to be the hard part. Fable 5.1 failed at it miserably and the reason is really cool! So what is the Shadow Test? The rules actually fit in one sentence: build one object whose three shadows spell a word. It's the trick from the cover of Gödel, Escher, Bach. The models get a 10×10×10 grid, three lights and a target word. They answer with nothing but a list of cube coordinates. No tools, one shot, and a script does the scoring, not a human. The first version was too easy. Every model scored 100%, because they all know the textbook trick. So I added one rule: every cube must be glued to the table through its neighbours. Anything floating falls off. The textbook answer went from 100% to 0%, immediately. Then I used an exact solver to prove the ceiling. For "AGI", the best possible score is 92.4%. That makes "optimal" a proof, not an opinion. 6 frontier models, 5 tries each. Results: 🥇 GPT-6 Astra: 92.4%, the fewest cubes (78) 🥈 Opus 5.5: 92.4% 🥉 Sonnet 5.5: 92.4% Gemini 3.8 Flash: 89.9% GPT-6 Sol: 88.9% Fable 5.1: 45.5% Only three models hit the proven optimum on every try. So as you can see, Level 1 is already solved, and that's by design. But the Shadow Test is a ladder: -> L2: a 16³ grid -> L3: a fourth light at 45° -> L4: no glue, stacking only The question isn't "does it pass". It's how high it climbs. Every sculpture from every model, including Fable's collapse, is below in 3D. Turn gravity off to see what it thought it built 👇 Also, reply with 3 letters. The best reply becomes the next episode!