Recent releases of OpenAI's GPT-6 series in September 2026 have lifted the MathArena expected performance frontier to 90.1% for GPT-6.1 Sol (max), up from the 63-74% range achieved by GPT-5.5 earlier in the year. MathArena aggregates results across uncontaminated final-answer, proof, and research-level math competitions released after model cutoffs, with average scores across tracked models rising from 45% to 84% over the past twelve months. Anthropic's Claude Opus 5.5 and Fable variants sit just behind at 82-83%, while open models lag further. Continued scaling, post-training reasoning improvements, and potential new model drops before year-end remain the main swing factors for whether any system crosses higher thresholds on this evolving benchmark.