GPT-6.2 Astra Tops MazeBench as Costs Drop; Gemini 4 Neon Leads in Token Efficiency

On the MazeBench 3D open-world benchmark with a uniform 200M token budget, GPT-6.2 Astra leads with 16% while cutting per-run cost from $840 to $225; GPT-6.2 Sol nears Astra at similar price, and Gemini 4 Neon shows far superior uncached token efficiency.

2026-11-20 ~ 2026-11-20 · 4 related posts