BALROG: leaderboard
Measures ability to play video games. Tests LLMs and VLMs as autonomous agents across game environments (BabyAI, Crafter, NetHack, etc.). Measures planning, spatial reasoning, and exploration.
Metric: Progress (%). Source: balrogai.com. 39 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 100 |
| 2 | Claude Opus 5 (Max) | 100 |
| 3 | GPT-6 Astra (Max) | 100 |
| 4 | GPT-5.6 Sol (Max) | 98 |
| 5 | GPT-5.6 Luna (Max) | 98 |
| 6 | GPT-5.6 Terra (Max) | 98 |
| 7 | Gemini 3.1 Pro (Preview) (Thinking) | 98 |
| 8 | Gemini 3 Pro | 96 |
| 9 | Gemini 3 Flash | 86 |
| 10 | Gemini 2.5 Pro (Preview 03-25) | 80 |
| 11 | Claude Opus 4.5 | 80 |
| 12 | GPT-5 (Minimal) | 80 |
| 13 | GPT-4o (2024-05-13) | 77.6 |
| 14 | Grok 4 | 76 |
| 15 | Reka Flash 3 | 76 |
Interactive version: theaggregate.ai/benchmark?slug=balrog-w48nbo · How It Works · Data refreshed daily, snapshot 2026-10-09.