Can a language model solve a Rubik's Cube in one reply?

Each move is a letter for a face, and an optional mark for the direction.

• The letter names the face to turn: U up, D down, L left, R right, F front, B back.
• No mark means a quarter turn clockwise, as you look straight at that face.
• An apostrophe (R') means a quarter turn anticlockwise.
• A 2 (R2) means a half turn.

That gives 18 moves in all:

So R U R' U' means: right face clockwise, top clockwise, right face back, top back. Every answer on this site is a list like that. Our engine plays it on the cube and checks the result.

A model sees a scrambled cube as text: all 54 stickers on an unfolded map. It must reply with one list of moves that solves it.

• One reply. No tools, no feedback, no second try. Ask ChatGPT or Claude in their apps and they write a Python script and solve it every time. That is not what this measures.
• The same cubes for every model. The cubes come from fixed seeds, so every model sees the same cubes in the same order.
• Checked by the engine. Our code plays the moves. A cube counts as solved only when all 54 stickers are back in place.
• 3 moves is an entry check, not a score. A model gets up to 5 cubes scrambled with 3 moves. One solve is a pass, because nobody writes the exact 3-move answer by luck. A fail means it solved none of the 5. We nearly dropped Claude Sonnet 5 after one failed cube, which is why one failure is not enough.
• 6, 10 and 20 moves are scored, on 10 different cubes each. A model must pass one length to go on to the next. The scale stops at 20 because of God's number: every position can be solved in 20 moves or fewer, so a 100-move scramble is no harder than a 20-move one.

The full prompt is on the leaderboard, under Method. Every model page shows the prompt, the answer and a replay for each cube.

31 of the 37 models that took the entry check passed it. Only 11 of them solved even one cube scrambled with 6 moves. The leaderboard has the current order.

The models that succeed write long answers. Past 3 moves they do not look for the shortest route. They solve the cube the way people learn to: one layer at a time, with fixed move sequences. It is slow but safe, because each step fixes one piece and leaves the solved pieces alone. GPT-6 Astra averages about 74 moves for a cube that is about 7 moves from solved. When we asked GPT-5.6 Sol for 20 moves or fewer, it wrote exactly 20 and failed. The normal prompt on the same cube got a 151-move answer that worked.

The task is not out of reach for a neural network. A small model we trained ourselves, with a search on top, solves 24 of 25 cubes scrambled with 20 moves. It is on its own page: Tiny Cube Solver.

Many models let you set how hard they think, from low to max. That thinking comes out of the same token limit as the answer, so more thinking leaves less room to write it.

• OpenAI models improve with more thinking. Sol went from 1/10 at medium to 7/10 at xhigh.
• Anthropic models get worse. On the same 10 cubes, Claude Opus 5.5 solved 5 at low, 3 at high and 2 at max. At high settings they often used the whole limit and wrote nothing. Claude Fable 5.1 at max solved 2 of 10 and wrote nothing on the other 8, at $6.72 a call.

So each model starts at its default setting, and we move it up or down depending on how it failed.

Providers do not return the same answer every time, even when we ask for the most predictable one. One model solved 0 of 10 cubes, then 2 of the same 10 on a second run. So read any score out of 10 as give or take 2.