Recent Anthropic Claude releases, particularly Fable 5.2 and Opus 5.5, have posted the highest HLE scores across multiple leaderboards at 65–74% depending on tool use and protocol, outpacing OpenAI’s GPT-6 Astra at roughly 54–57% on the same evaluations. HLE’s 2,500 expert-authored questions test frontier reasoning and knowledge with minimal leakage, and variations in tool access or verification explain score spreads. OpenAI’s latest models demonstrate competitive scaling on reasoning subsets yet trail in overall accuracy, reflecting differences in training emphasis and post-training. Additional GPT variants or updates before year-end remain possible catalysts, though current verified results place OpenAI behind the top non-OpenAI systems on this benchmark.
基于Polymarket数据的AI实验性摘要。这不是交易建议,也不影响该市场的结算方式。 · 更新于1H6H1D7D30DALL
View resolved

警惕外部链接哦。
警惕外部链接哦。
常见问题