BALROG MiniHack (VLM): leaderboard
Vision-language models on MiniHack dungeon tasks, measuring visual planning, inventory management, and survival in procedurally generated environments.
Metric: Progress (%). Source: balrogai.com. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 22.5 | #291 |
| 2 | Gemini 2.5 Pro (Preview 03-25) | 12.5 | #123 |
| 3 | Gemini 1.5 Pro (002) | 5 | #349 |
| 4 | GPT-4o (2024-05-13) | 5 | #325 |
| 5 | Qwen 2 VL 7B Instruct | 5 | #826 |
| 6 | Gemini 1.5 Flash (002) | 2.5 | #583 |
| 7 | GPT-4o Mini (2024-07-18) | 2.5 | #558 |
| 8 | Llama 3.2 11B Instruct | 2.5 | #1126 |
| 9 | Llama 3.2 90B Instruct | 2.5 | #601 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=balrog-minihack-vlm · How It Works · Data refreshed daily, snapshot 2026-10-20.