Claude Opus gains 250 Elo playing 200 chess games, while GPT-6 Astra drops 410

新智元 · wechat · 2026-10-05

Peter Gostev, AI capability lead at Arena, ran an experiment letting models play 200 chess games against Stockfish with no weight updates — only self-written notes for memory.

The experiment probes whether frozen-weight models can self-improve through in-context notes — Opus's +250 suggests yes.

Related event: AI Continuous Learning Test: Opus Gains Elo Over 200 Chess Games While GPT Regresses(3 posts)→

Original post →

More from Models

Models channel →