Multi-turn Debate (Lechmazur) — leaderboard
Adversarial multi-turn debate benchmark where LLMs argue opposing sides of policy propositions. Each matchup runs twice with sides reversed. Ranked via Bradley-Terry ratings across 1,000+ debates judged by LLMs. Tests factual accuracy under pressure, counterargument quality, and epistemic discipline.
Metric: Bradley-Terry Rating. Source: github.com. Status: saturation imminent. 41 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5 (High) | 1757.2 |
| 2 | Kimi K3 | 1740.8 |
| 3 | Claude Opus 4.7 (High) | 1684.7 |
| 4 | GPT-5.6 Sol (High) | 1680.5 |
| 5 | Claude Opus 4.8 (High) | 1668 |
| 6 | Claude Sonnet 5 (High) | 1618.5 |
| 7 | Claude Sonnet 4.6 (High) | 1597.2 |
| 8 | GLM-5.2 (Max) | 1594.9 |
| 9 | GPT-5.4 (High) | 1583.7 |
| 10 | GPT-5.5 (High) | 1563.6 |
| 11 | GLM-5.1 | 1554.7 |
| 12 | Kimi K2.6 | 1547.4 |
| 13 | MiniMax-M3 | 1537.4 |
| 14 | GPT-5.4 (Non-reasoning) | 1531.4 |
| 15 | Gemini 3.1 Pro (Preview) | 1529.1 |
Interactive version: theaggregate.ai/benchmark?slug=multi-turn-debate-lechmazur · How the rankings work · Data refreshed daily, snapshot 2026-07-22.