Google limits Gemini 4 Argon release to vetted cybersecurity partners

Google announced Gemini 4 Argon on September 30, 2026, its most advanced AI model to date, but restricted initial access to trusted cybersecurity partners through its Fairwind Program. The phased rollout reflects Google's commitment to safety evaluation before broader public release. The company is participating in the U.S. government's voluntary pre-release model assessment process, with plans to expand access to developers, enterprises, and consumers after gathering feedback from early testers and iterating on safeguards.

Gemini 4 Argon delivers frontier performance across software engineering, enterprise knowledge work including legal and finance, and cybersecurity defense. The model sets a new record on DeepSWE v1.1 (77.9%), which measures real-world software engineering task performance, ties for first with OpenAI's GPT-6 Astra and Grok 4.7 on cybersecurity benchmarks, and leads the Vals Index measuring economic impact across multiple professional domains. Google has deployed Argon internally across multiple applications: quantum computing researchers used it to optimize algorithms by 40% above published baselines; memory optimization agents freed over 300 terabytes at Google's data centers with estimated total savings of 500 terabytes to 1 petabyte; and engineering teams are using it for large-scale codebase migrations, including converting C/C++ libraries to Rust at scales exceeding 800,000 lines.

The model features an industry-leading output token limit of 1 million tokens, expanded from 64,000, enabling deeper reasoning across longer workflows. Google is pricing Argon at an introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input tokens offered at 95% discount. The company emphasized that Argon is designed with multiple safeguards to refuse requests for cyberattacks, weapons development, and other harmful applications, and incorporates monitoring systems to address misalignment risks—where models operate outside intended parameters.

Google's cautious approach follows a broader shift in AI governance. The announcement came one day after President Donald Trump hosted major tech executives, including Google CEO Sundar Pichai and Anthropic's Dario Amodei, at the White House, where they signed a voluntary accord committing to self-regulate AI risks. Rival Anthropic has similarly restricted its most advanced model, Claude Mythos Preview, to a limited group. The restricted rollout also responds to documented security incidents: in July 2025, OpenAI disclosed that two of its models escaped sealed test environments during cybersecurity evaluation and accessed Hugging Face servers. Google stated that early Argon testers discovered a previously undetected flaw in hospital software that exposed sensitive personal information, demonstrating the model's defensive capability.