An aerial view of Google’s data center outside Pryor, Oklahoma, photographed in June 2019. Archival infrastructure photograph: Xpda / Wikimedia Commons, CC BY-SA 4.0.
Google has announced Gemini 4 Argon, a new AI model aimed at demanding software and professional work, but its initial availability is narrower than a typical consumer product launch. The September 30 announcement says access is beginning with selected cybersecurity defenders through Google DeepMind’s Fairwind Program.
Google is emphasizing the model’s ability to sustain lengthy, multi-step work. Its advertised output limit rises to one million tokens, compared with the previous 64,000-token limit cited in the announcement. That measures how much the model can generate during a run; it should not be confused with an input-context specification or a guarantee that longer answers are better.
Who can use it first
The Fairwind Program’s published rules give priority to governments, critical infrastructure operators and core technology platforms. Applicants are vetted, and approved organizations must restrict access to relevant internal security teams. The program does not allow participants to resell or redistribute access.
Fairwind partners can use Argon alongside CodeMender, Google’s code-security agent, to help investigate vulnerabilities and prepare fixes. The program permits authorized defensive and academic work, including threat simulation and malware analysis. These access conditions are important context for claims about autonomous cybersecurity capabilities: the early deployment comes with organizational controls and a defined purpose.
What broader availability will mean
Google says broader access will begin with paid API customers and Google AI Ultra subscribers, after further testing and safeguards work. It has not specified a public release date in the announcement. Introductory API pricing is listed at $2 per million input tokens and $10 per million output tokens; the stated post-introductory rates are $4 and $20 respectively.
The practical takeaway for developers is to separate a model announcement from a service they can build into a production workflow today. Access terms, final pricing and performance on a team’s own tasks can matter as much as benchmark results.
For consumers, the immediate significance is the direction of development: AI systems are being presented as tools for longer assignments that involve repeated reasoning and action. Whether that delivers dependable value outside controlled tests will require evidence from broader use. For now, Argon’s rollout remains a staged release, with the first users selected for security work.
