DIY benchmark pits 14 LLMs vs 27 humans in 2,038 Cambridge Union debates, Fable 5.2 tops Elo

eliebakouch · x · 2026-11-18

A self-built benchmark replays real Cambridge Union motions with 14 LLMs and 27 humans debating head-to-head — 2,038 debates total, judged anonymously by an 8-LLM jury with audience questions from other models.

Elo results: Claude Fable 5.2 leads at 1821 (259–16), followed by GPT-6.2 Astra at 1715, then Fable 5.1 and Opus 5.5. Humans Paul Graham (1245) and Henry Shevlin (1158) land below GLM 5.3 but above Gemini 4 Hydrogen; Mistral Large 4 is last at 941 (25–236). Top models now clearly out-debate the human participants.

Original post →

More from Fun

Fun channel →