GT-HarmBench: leaderboard
Metric: Utilitarian accuracy (%): share of the 1,535 scenarios in which the pair of choices maximizes total welfare (sum of both players' utilities), averaged over the six game types weighted by their scenario counts; self-play: two copies of the model each read their side of a natural-language high-stakes scenario built from the MIT AI Risk Repository and choose an action independently, with no communication; reasoning at medium where available; a propensity to choose the socially optimal joint action, not task accuracy; higher is better. Source: arxiv.org. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.3 70B Instruct | 81 | #520 |
| 2 | GPT-4o | 75 | #333 |
| 3 | GPT-5 Mini | 64 | #176 |
| 4 | GPT-5.1 (Medium) | 64 | #131 (GPT-5.1) |
| 5 | Gemini 3 Flash (Medium) | 64 | #93 (Gemini 3 Flash) |
| 6 | GPT-5.2 (Medium) | 63 | #105 (GPT-5.2) |
| 7 | GPT-5 Nano | 59 | #415 |
| 8 | Qwen 3 8B (Thinking) | 48 | #667 (Qwen 3 8B) |
| 9 | DeepSeek V3.2 (Thinking) | 46 | #198 (DeepSeek V3.2) |
| 10 | Grok 4.1 Fast (Reasoning) | 35 | #208 (Grok 4.1 Fast) |
| 11 | Qwen 3 30B A3B (Thinking) | 33 | #488 (Qwen 3 30B A3B) |
Interactive version: theaggregate.ai/benchmark?slug=gt-harmbench · How It Works · Data refreshed daily, snapshot 2026-10-11.