Google announced Gemini 4 Argon on September 30, 2026, positioning the frontier model for deep reasoning in long-horizon tasks like software engineering, enterprise knowledge work, and cybersecurity defense. It introduces a 1-million-token output limit and leads several internal benchmarks such as DeepSWE v1.1 at 77.9%, though it remains in limited rollout via the Fairwind Program for vetted cyber defenders and Google teams. On Humanity’s Last Exam (HLE), prior Gemini versions reached the mid-40s percent range without tools, while current leaders like Claude Fable 5.1 top recent leaderboards near 65 percent with tools; no Argon HLE results have surfaced yet. Traders are watching for initial public evaluations or API access that could enable benchmark testing, alongside ongoing competition from OpenAI and Anthropic models advancing on the expert-level academic questions.