BALROG Official LLM Leaderboard: leaderboard
Official BALROG LLM leaderboard for long-horizon game reasoning across BabyAI, Crafter, BabaIsAI, MiniHack, NetHack, and related game environments.
Metric: % Progress (percentage). Source: github.com. Source: benchmarklist.com. Saturation forecast: Around July 2027. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 58.1 | #80 |
| 2 | Gemini 3.1 Pro (Preview) (Thinking) | 57 | #58 (Gemini 3.1 Pro (Preview)) |
| 3 | Gemini 3.1 Pro (Preview) | 56.9 | #58 |
| 4 | Gemini 3 Flash | 48.1 | #99 |
| 5 | Grok 4 | 43.6 | #181 |
| 6 | Claude Opus 4.5 | 43.5 | #83 |
| 7 | Gemini 2.5 Pro (Preview 03-25) | 43.3 | #123 |
| 8 | Claude Opus 4.5 (Thinking) | 43 | #83 (Claude Opus 4.5) |
| 9 | DeepSeek R1 | 34.9 | #251 |
| 10 | Gemini 2.5 Flash | 33.5 | #245 |
| 11 | GPT-5 (Minimal) | 32.8 | #98 (GPT-5) |
| 12 | Claude 3.5 Sonnet (20241022) | 32.6 | #291 |
| 13 | GPT-4o (2024-05-13) | 32.3 | #325 |
| 14 | Claude Haiku 4.5 | 31.2 | #280 |
| 15 | Grok 3 | 29.5 | #304 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=balrog-official-llm-leaderboard · How It Works · Data refreshed daily, snapshot 2026-10-20.