Recent releases of OpenAI’s GPT-6 Astra in early September and Anthropic’s Claude-Fable-5.1 have driven top MathArena expected performance to 88% and 84%, respectively, following rapid gains that lifted FrontierMath Tier 4 accuracy from roughly 5% to nearly 98% in 14 months. These models lead on research-level ArXivMath, BrokenArXiv, and proof-based tracks, where benchmark maintainers have already introduced harder problems and revised grading to combat saturation. Open-weight contenders such as Qwen3.8-Max sit near 56%, widening the gap while highlighting competitive pressure from closed labs. With only three months remaining until year-end, traders are watching for additional frontier releases, further benchmark refreshes, and any demonstration of sustained progress on the most demanding math tasks.