Google's Gemini 4 Argon Tops Benchmark Race, Reclaiming Frontier AI Lead with 1M Token Output

Google has announced Gemini 4 Argon, a new frontier model that outperforms OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 on key software engineering benchmarks, marking a significant shift in the competitive landscape of advanced AI systems. The model, unveiled September 30, 2026, achieved 77.9% on DeepSWE v1.1—a test measuring real-world long-horizon software engineering tasks—compared with Opus 5.5's 74.2% and Astra's 74.1%. On the security-focused CWE-bench v1, Argon tied for first place at 68%.