GPT-6 models solve 86% of a brand-new puzzle game, study finds
A report by Gavin Leech and colleagues finds GPT-6 Astra and GPT-6.1 Sol solve about 86% of a difficult puzzle game released after their training cutoff, suggesting game-based benchmarks for frontier models are losing effectiveness.
2026-11-17 ~ 2026-11-17 · 2 related posts
- Frontier models ~solve a brand-new game, say researchers measuring generalisation is dying — gleech · 2026-11-17
- Frontier models solve 86% of a brand-new puzzle game — videogame evals are dead — gleech · 2026-11-17