Recent releases from Anthropic, including Claude Opus 5 and Mythos variants, have driven top MathArena scores above 84% on uncontaminated competition problems as of mid-2026, reflecting advances in extended reasoning chains and tool use that outpace prior GPT-5 iterations. OpenAI and Moonshot models follow closely, with competitive dynamics centered on scaling inference-time compute and synthetic math data. Traders monitor upcoming frontier releases expected before year-end, alongside benchmarks like FrontierMath tiers, where current capabilities remain below 50% on harder sets. These developments create strong but not certain momentum toward higher thresholds by December 31, tempered by the risk of incremental rather than breakthrough gains.