Recent leaderboards show Anthropic’s Claude Fable 5.2 topping standard Humanity’s Last Exam evaluations at 74%, with OpenAI’s GPT-6 Astra reported around 55-60% on Scale, Artificial Analysis, and tool-augmented variants as of mid-November 2026. This gap reflects differences in post-training emphasis and benchmark protocols, where tool access or verification subsets can shift scores by 10+ points. The 2,500-question expert benchmark, created by the Center for AI Safety and Scale AI, tests frontier multidisciplinary reasoning and has seen roughly 30-point gains across labs in the past year. With the market resolving on the official HLE leaderboard by year-end, any unreleased OpenAI model or update before December 31 could alter the highest achieved score, though current trajectories favor modest gains over dramatic jumps.
Résumé expérimental généré par IA à partir des données Polymarket. Ceci n'est pas un conseil de trading et ne joue aucun rôle dans la résolution de ce marché. · Mis à jour1H6H1D7D30DALL
View resolved

Méfiez-vous des liens externes.
Méfiez-vous des liens externes.
Questions fréquentes