Debate Benchmark adds 8 models: Qwen 4 Max jumps 144 points, Claude Fable 5.2 leads at 1688
zero0_one1 · reddit · 2026-11-19
Lech Mazur's Debate Benchmark update adds 8 models. The benchmark tests whether a model can defend a position under sustained opposition — answering objections, asking pointed questions, and closing well — now covering 10,928 rated debates across 683 topics with 33,342 judge decisions. Each matchup runs twice with PRO/CON swapped, judged independently by three models from different families.
Key scores: Claude Fable 5.2 tops new entries at 1688 (18 below Fable 5.1); Qwen 4 Max posts the biggest gain, 1535 → 1679 (+144); Gemini 4 Carbon hits 1676 (+27 over Argon); GPT-6.2 Astra 1673 (+67); Tencent Hy4 (high) +85 to 1646; Step 5 Preview debuts at 1521; MiniMax M3.1 Flash 1509; Seed 2.1 Pro 1468. Law/regulation is the largest topic group (135 motions).
More from Models
- Free Gemini users see model picker shrink to Auto-only within a single month — python_buddy · 2026-11-19
- Google quietly rewrote Gemini 4 Neon docs, dropping "Flash-tier cost" disclosure — lyraxana · 2026-11-19
- Gemini 4 Neon High matches xHigh scores at 42% lower cost — adonis_singh · 2026-11-19
- Google shipped Argon, Carbon, Hydrogen and Neon in 5 weeks of Gemini 4 iteration — Afinetheorem · 2026-11-19
- GPT 6.2 Astra shows surprisingly strong zero-shot crystal generation, a year after the task looked hopeless — xennygrimmato_ · 2026-11-19
- LLMs zero-shot beat fine-tuned world action models on 7 of 10 robot tasks — paigeinsf · 2026-11-19