MazeBench: leaderboard
MazeBench evaluates long-running agents in a visual-spatial environment where they navigate three-dimensional rooms, solve Sokoban-style puzzles, and collect hidden gems. It supports vision, ASCII, and JSON observation modes, with separate tracks for agents with and without Python tool access. The pinned metric is Gems collected, measuring gem collection in the leaderboard results. Source: BenchmarkList’s attributed result table.
Metric: Gems collected (gems). Source: benchmarklist.com. Saturation forecast: Around February 2028. 44 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 14 | #6 (GPT-6 Astra) |
| 2 | GPT-5.6 Sol (Max) | 13 | #22 (GPT-5.6 Sol) |
| 3 | Claude Opus 5 (Max) | 12 | #16 (Claude Opus 5) |
| 4 | Claude Fable 5 (Max) | 11 | #20 (Claude Fable 5) |
| 5 | Claude Fable 5.1 (Max) | 10 | #10 (Claude Fable 5.1) |
| 6 | GPT-5.6 Terra (Max) | 6 | #40 (GPT-5.6 Terra) |
| 7 | Gemini 3.8 Flash (High) | 4 | #30 (Gemini 3.8 Flash) |
| 8 | Kimi K3 (Max) | 1 | #35 (Kimi K3) |
| 9 | Gemini 3.1 Pro (Preview) (High) | 1 | #64 (Gemini 3.1 Pro (Preview)) |
| 10 | Gemini 3.7 Flash (High) | 1 | #24 (Gemini 3.7 Flash) |
| 11 | Grok 4.6 (xHigh) | 1 | #29 (Grok 4.6) |
| 12 | Claude Haiku 4.5 | 0 | #287 |
| 13 | Llama 4 Maverick | 0 | #486 |
| 14 | GPT-5.5 (xHigh) | 0 | #33 (GPT-5.5) |
| 15 | GPT-5.6 Luna (Max) | 0 | #79 (GPT-5.6 Luna) |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=mazebench · How It Works · Data refreshed daily, snapshot 2026-11-05.