Google announced Gemini 4 Argon on September 30, 2026, as its latest frontier large language model optimized for long-horizon reasoning in software engineering, enterprise knowledge work, and cybersecurity. The model introduces a 1-million-token output limit and leads several agentic coding and knowledge benchmarks such as DeepSWE v1.1 at 77.9 percent and the Vals Index, though it has not yet posted public results on Humanity’s Last Exam (HLE). Current HLE leaderboards show Claude Fable 5.1 and Opus 5 variants ahead at 64–65 percent without tools, with GPT-6 variants close behind; earlier Gemini iterations reached the mid-40s percent range by mid-2026. Traders are watching for official HLE evaluations or internal progress reports that could clarify Argon’s standing ahead of broader API rollout.