Frontier AI labs continue rapid iteration on mathematical reasoning, with OpenAI’s GPT-6 series (including GPT-6.1 Sol) posting the highest MathArena expected performance near 90% in late September 2026 releases. Anthropic’s Claude Opus and Fable variants, alongside Moonshot’s Kimi K2.6 and Z.AI’s GLM-5 models, cluster closely behind on uncontaminated olympiad-style benchmarks such as AIME, USAMO, and FrontierMath. Traders are watching for further capability jumps from scheduled updates or new training runs before year-end, as saturation on easier MATH and AIME sets pushes evaluation toward harder proof and research-level tasks. Recent full IMO scores by multiple systems underscore the pace, though exact MathArena thresholds remain sensitive to prompt strategies and evaluation protocols.