GPT-6 Astra struggles on MazeBench 3D spatial reasoning benchmark

Blogger patiencecave tested GPT-6 Astra on MazeBench, a self-built 3D open-world spatial reasoning benchmark: the model ran for over 60 hours on the evaluation and scored only 14%. While low in absolute terms, this is the strongest performance on MazeBench to date—and achieved without code execution, beating GPT-5.6 Sol's 13% even with Python code execution enabled—leading the blogger to see it as a generational leap in capability.

Confirmed

• To save tokens, the benchmark allowed GPT-6 Astra to submit moves in batches each turn rather than step by step; the model then consistently planned 10-20 moves ahead, cutting roughly 3 billion tokens of consumption down to 350 million.
• Astra outperformed all previous agents, easily solving puzzles requiring 5 moving parts to be handled simultaneously; exploration heatmaps show its exploration range far exceeding the comparison model Fable 5.1.
• Failure mode analysis shows Astra tends to interpret levels as top-down 2D planes, treating walls only visible from a 3D perspective as invisible barriers; it can complete tutorial levels but fails 3D puzzles where objects are stacked more than 3 layers high.

Why it matters

Even the strongest new-generation models still have structural weaknesses in long-horizon, true 3D spatial reasoning, suggesting that "top-down-view" world understanding remains a bottleneck for current models; meanwhile, the sharp drop in token consumption from batch planning also demonstrates how agent benchmark design affects cost. The blogger notes there is still a large gap to passing MazeBench's harder challenges.

2026-09-07 ~ 2026-09-07 · 7 related posts

• [source] GPT-6 'Astra' burns 60+ hours in 3D spatial reasoning eval, scores just 14% — patience_cave · 2026-09-07
• GPT-6 Astra outperforms all prior agents, juggling 5 moving puzzle parts — patience_cave · 2026-09-07
• GPT-6 treats 3D levels as top-down maps, failing puzzles stacked over 3 blocks — patience_cave · 2026-09-07
• Batched moves let GPT-6 plan 10-20 steps ahead, cutting 3B tokens to 350M — patience_cave · 2026-09-07
• [source] GPT-6 without code execution beats GPT-5.6 with Python on MazeBench — patience_cave · 2026-09-07
• [source] GPT Astra scores just 13% on MazeBench without tools — Wonderful_Buffalo_32 · 2026-09-07

1 near-duplicate retellings: basedjensen