Google announced Gemini 4 Argon on September 30, 2026, positioning it as its new frontier large language model optimized for long-horizon software engineering, enterprise tasks in legal and finance, and cybersecurity defense. The model features an expanded 1-million-token output limit and leads or ties most internal benchmarks against GPT-6 Astra and Claude Opus 5.5, including a 77.9% score on DeepSWE v1.1. It is currently limited to vetted participants in the Fairwind Program and Google’s internal teams while the company completes U.S. government voluntary safety reviews; broader API and consumer access is planned next for paid users and Google AI Ultra subscribers but carries no announced timeline. This phased approach, driven by dual-use risks in cyber defense, creates the main uncertainty for release-date markets.