LLM Rubik's Cube Race

Four models get the same scrambled 3x3 cube and race to solve it. Each panel shows the model's moves on a 3D cube with time, token usage and cost. Replay plays a recorded race with its original timing. Live runs call the models and need the access key.

How it works

• Claude Fable 5.1, GPT-6 Astra and Grok 4.6 get the cube as a text net and reply once with a full move sequence. Their moves are applied to the cube after the reply arrives.
• Jev is a TypeSafe decision model, not a text generator. It gets the cube state and 18 candidate turns, picks one, and is asked again with the new state until the cube is solved or the step limit is reached.
• All four run through OpenRouter. Cost comes from OpenRouter's usage report for each request.
• A replay is the event log of a real live race, played back at the recorded pace. Nothing is simulated and nothing is re-run. The speed control only changes playback, not the recorded times.
• A model counts as solved only when the moves it produced, applied to the scramble, leave every face a single color. Any moves outside U D R L F B notation are ignored.

Reading the numbers

• Time is wall-clock from request start to the last token (LLMs) or the last decision (Jev).
• Tokens are input / output as reported by OpenRouter. Reasoning tokens are included in output.
• Scramble depth is the number of random turns applied to a solved cube. 20 is a full scramble. Depth 1 to 3 is where text models have a real chance.
• While a text model works, its reasoning and reply stream into the panel as they arrive. Some providers send reasoning only as short summaries, so the text can pause for a while. The cube only turns once the reply is in.
• Every model is cut off at the time limit. A timed-out model still gets any moves found in its partial reply, and its tokens and cost come from OpenRouter's record of the generation.