Anthropic’s September 2026 releases of Claude Opus 5.5, Sonnet 5.5, and related Fable/Mythos 5.1 variants have propelled the company’s models to the top of multiple Humanity’s Last Exam leaderboards, with scores reaching 61–67% on tool-augmented runs and 55–65% on standard evaluations. These gains stem from targeted advances in long-horizon reasoning, tool use, and agentic coding that align directly with HLE’s expert-level questions across math, science, and humanities. OpenAI’s GPT-6 Astra and Google’s Gemini 4 models trail but remain competitive, while rapid iteration by Anthropic and benchmark refinements like HLE-Diamond keep the ceiling uncertain. With three months left in 2026, further Claude updates could lift the highest score toward or beyond 65%, though saturation risks and evaluation variances remain key variables for traders.