ParliamentBench: leaderboard
Overall win rate across 100 Secret Hitler games against Llama 3.3 70B Instruct opponents. Source-local instruction checkpoints are retained; reasoning uses low or comparable settings where available. Human, random and algorithmic reference rows are not model scores.
Metric: Overall Win Rate (%). Source: arxiv.org. Saturation forecast: Around January 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 81 | #79 |
| 2 | Kimi K2.5 | 76 | #144 |
| 3 | Grok 4.1 Fast | 69 | #217 |
| 4 | DeepSeek V3.1 Terminus | 66 | #220 |
| 5 | Llama 3.3 70B Instruct | 58 | #534 |
| 6 | Qwen 3.5 397B A17B | 56 | #146 |
| 7 | Olmo 3.1 32B Instruct | 53 | #763 |
| 8 | GPT-OSS-120B | 50 | #341 |
| 9 | DeepSeek R1 | 46 | #251 |
| 10 | Mistral Small 3 | 43 | #627 |
| 11 | Llama 3.1 70B Instruct | 42 | #561 |
| 12 | Gemma 3 27B (IT) | 39 | #523 |
| 13 | GPT-OSS-20B | 24 | #518 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=parliamentbench · How It Works · Data refreshed daily, snapshot 2026-10-20.