Recent releases of frontier large language models have accelerated progress on the MathArena benchmark, which evaluates uncontaminated math competition problems including AIME, HMMT, and proof-based contests. OpenAI's GPT-6 Astra, added in early September 2026, leads with expected performance near 90.7 percent across recent evaluations, ahead of Anthropic's Claude Opus 5 and Fable 5.1 series at roughly 70-72 percent. Moonshot AI's Kimi models and Alibaba's Qwen variants also post strong results on weighted math metrics, reflecting gains in reasoning chains and generalization. With multiple labs iterating rapidly through the fall, trader sentiment hinges on whether incremental improvements or a major release push any model past the market's implied threshold by year-end.