Recent releases of frontier models have driven gains on math benchmarks like MathArena, where Anthropic’s Claude-Opus-5 (max) reached 84.4% overall in late July 2026, edging out OpenAI’s GPT-5.6 variants near 80%. Traders focus on continued scaling of reasoning chains, test-time compute, and post-training on harder problems such as FrontierMath and ArXiv-sourced contests, which remain below 60% for even the strongest systems. Competitive pressure from Google, xAI, and Chinese labs like Moonshot and Qwen, plus expected model drops through Q4, shapes sentiment around whether any system can clear the market’s implied threshold by year-end.