MazeBench maze eval: flagship models score 0%, Gemini 3.7 Flash leads

In late August, MazeBench published a new round of Terra max maze evaluations: ox-alpha, Grok 4.6, GLM-5.3 and Qwen 3.8-max all scored 0%, while Gemini 3.7 Flash led with 1%, ahead of Kimi K3. The results point to a clear weakness in spatial-reasoning maze tasks among frontier models.

Confirmed

• In MazeBench's Terra max test, ox-alpha, Grok 4.6, GLM-5.3 and Qwen 3.8-max all scored 0%
• Gemini 3.7 Flash scored 1%, outperforming Kimi K3
• Per @patiencecave's coding-agent breakdown: Grok 4.6 explored the most rooms (6) yet scored 0% on puzzle solving; Gemini 3.7 Flash was the only model to solve the puzzle (collecting gems)

Why it matters

• Multiple flagship models scoring 0% on the same task shows maze-style exploration and spatial planning remain a blind spot for frontier models
• The contrast between Grok 4.6's broad exploration with zero solving and Gemini's task completion suggests exploration breadth and task completion can decouple, offering a new lens for model evaluation

2026-08-24 ~ 2026-08-26 · 4 related posts

• [source] MazeBench results: flagship models score 0% on maze tasks — patience_cave · 2026-08-24
• [source] MazeBench Update: Grok explores most, Gemini Flash solves puzzles while others struggle with T-shaped blocks — patience_cave · 2026-08-25
• MazeBench Results: Grok 4.6 and Peers Score Near 0% — Artemonim · 2026-08-25
• [source] MazeBench: Top Models Score 0%, Gemini Flash Beats Kimi — maziarmazi23572 · 2026-08-26