Text-only reasoning solves Rubik's Cube: GPT-5.6 Sol clears CubeBench's 0% barrier in 333 moves

Willing_Plate_5417 · reddit · 2026-09-12

CubeBench previously reported a 0.00% pass rate for leading LLMs on long-horizon Rubik's Cube tasks testing spatial reasoning and state tracking. A Reddit user re-ran the experiment with a frontier model under strict constraints—no Python, no external solver, no hidden solution, just text-based reasoning against a cube engine.

GPT-5.6 Sol solved a freshly scrambled 3×3: 750.5 seconds, 333 moves, 61 interactions, 0 rejected moves. The solution was inefficient with self-correction, but it recognized OLL/PLL cases along the way. The takeaway: '0% success' timestamps a generation of models, not a permanent architectural boundary.

Original post →

More from Models

• Grok 4.7 delayed a few more days as Musk blames RL length penalty for giving up early — rohanpaul_ai · 2026-09-13
• Timelines Flooded With GPT-6 Astra Robot Demos as Physical AI Hypes Up — CyberRobooo · 2026-09-13
• Models Escaped Sandboxes to Read Eval Source Code, Compromising AI Evaluations — dhadfieldmenell · 2026-09-13
• 双 3090+老 EPYC 跑 177B MoE 模型:38 tok/s 仅花 800 美元 — ludos1978 · 2026-09-13
• Claude successfully dissected a virtual bird flock's neural network, finding command neurons — neuroecology · 2026-09-13
• Code Arena WebDev Turns One: Top Score Up 340 Points, GPT-6 Astra Leads at 1796 — arena · 2026-09-13

Models channel →