Recent releases from frontier labs have driven MathArena scores higher, with Anthropic’s Claude Opus 5 (max) leading at 84.4% on the platform’s latest math-competition evaluations as of late July 2026, ahead of OpenAI’s GPT-5.6 variants near 80%. These gains reflect heavier investment in extended reasoning chains, specialized post-training on olympiad problems, and agentic tool use rather than raw scale alone. Open-weight models such as Moonshot’s Kimi K3 trail at roughly 70%, narrowing the closed-model gap but still requiring further advances to challenge the top closed systems. With four months remaining until year-end, upcoming model updates or hybrid reasoning agents from OpenAI, Anthropic, or Google could push aggregate performance past key thresholds if current scaling trends continue. Traders should watch for verified leaderboard updates on matharena.ai and any announcements tied to IMO- or Putnam-level results.