MazeBench: GPT-6.2 Astra tops at 16% while costing 25% of its predecessor

patience_cave · x · 2026-11-20

Blogger patiencecave shares MazeBench results (a massive 3D open-world benchmark, all models given a 200M token budget): GPT-6.2 Astra scores 16%, GPT-6.2 Sol 14%, Gemini 4 Neon 10%. Astra matches GPT-6 Astra's intelligence per token; Neon ranks 5th on intelligence per token but leads decisively on price. Headline: GPT-6.2 Astra scores equivalently to GPT-6 Astra at 25% of the cost. Author notes fairness is hard to verify in 3D open-world evals.

Original post →

More from Models

Models channel →