OpenAI's GPT-6 Astra and GPT-6.1 Sol currently trail Anthropic's leading Claude Opus 5.5 and Fable 5.1 models on Humanity's Last Exam (HLE) leaderboards, posting scores near 54.7% versus 61%+ for the top closed models as of mid-October 2026. Rapid capability gains across frontier labs—driven by larger training runs, improved reasoning architectures, and tool-augmented evaluations—have lifted top HLE performance from single digits in early 2025 to the mid-60s range. OpenAI's early October release of 722 math manuscripts from an unreleased internal model highlights ongoing progress in expert-level reasoning that could lift scores before year-end. Key swing factors include any late-2026 GPT-6 variants, agentic fine-tunes, or verified high-compute runs that meet the benchmark's strict expert-question criteria.