Batched moves let GPT-6 plan 10-20 steps ahead, cutting 3B tokens to 350M

patience_cave · x · 2026-09-07

MazeBench allowed GPT-6 Astra to submit moves in batches instead of one per turn. The model consistently planned 10-20 moves ahead, collapsing 3 billion tokens into 350M. On the Klotski puzzle, batched Astra solved it 3x cheaper but took 60% longer to think.

Related event: GPT-6 Astra Sets MazeBench Record at 14% but Still Fails 3D Spatial Reasoning(5 posts)→

Original post →

More from Models

Models channel →