Recent releases from Anthropic and OpenAI, including Claude Opus 5 (max) at 84.4% and GPT-5.6 variants near 80% on MathArena's competition leaderboard, have driven trader sentiment toward yes. These models demonstrate continued gains on uncontaminated olympiad problems through scaled reasoning and post-training, though proof-heavy tasks lag below 40%. Moonshot's Kimi series and other open-weight entries trail at 60-70%, underscoring closed-lab advantages in math-specific capabilities. With four-plus months until year-end, labs are expected to iterate on agentic workflows and larger context windows; however, saturation on existing benchmarks and potential release delays introduce uncertainty around crossing higher thresholds like 90%.