Anthropic’s Claude Opus 5.5 and related variants currently lead Humanity’s Last Exam leaderboards at 61–67% accuracy on the core 2,500-question benchmark, with tool-augmented evaluations pushing some results higher. OpenAI’s GPT-6 Astra delivered a notable 82.9% on the refined HLE-Diamond subset in September 2026 when allowed web and code tools, illustrating how agentic capabilities can accelerate scores on this expert-level test spanning math, science, and humanities. Frontier labs continue releasing updated models monthly, with competitive pressure from Google Gemini variants and open-source entries like DeepSeek driving incremental gains. Traders focus on whether these trends sustain through December, as historical benchmark saturation patterns and remaining headroom on the original HLE suggest further upside remains possible before year-end.