MazeBench Exposes Weak 3D Spatial Reasoning in Frontier Models

A homemade 3D spatial reasoning benchmark, MazeBench, shows frontier models struggling: Opus 5.5 scored 6%, nearing 'Astra' level, while GPT-6 Sol and Grok 4.7 each scored just 1%.

2026-09-23 ~ 2026-09-23 · 2 related posts

• MazeBench 3D Spatial Reasoning: Opus 5.5 Hits 6% While GPT-6 Sol and Grok 4.7 Manage Just 1% — patience_cave · 2026-09-23

1 near-duplicate retellings: daniel_mac8