Google announced Gemini 4 Argon, its new frontier large language model, on September 30, 2026, positioning it for complex software engineering, enterprise knowledge work in legal and finance, and cybersecurity defense. The model features a 1 million-token output limit, up from 64,000 in prior versions, and leads benchmarks such as DeepSWE v1.1 at 77.9% while tying or exceeding rivals like OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 on coding and vulnerability tasks. Initial rollout is limited to vetted cyber defenders via Google’s Fairwind Program, with phased expansion planned for paid API users and Google AI Ultra subscribers after U.S. government safety evaluations. This follows Google’s decision to skip the earlier Gemini 3.5 Pro release due to performance shortfalls.