Google launches Gemini 4 Argon first to trusted cyber defenders

Google announced Gemini 4 Argon on September 30 and began rolling it out to trusted cyber defenders through the company’s Fairwind Program. [7] Google said releasing frontier capabilities at this level requires a phased approach and that it is participating in the U.S. government’s voluntary process for pre-release model access. [7]

The company plans to gather feedback from early testers and iterate on guardrails before making Argon available to developers, enterprises, and consumers. [7] VentureBeat reported that broader availability is planned to begin with paid API customers and Google AI Ultra subscribers. [9]

Google positions Argon for complex, long-horizon workflows spanning real-world software engineering, legal and financial knowledge work, and cybersecurity defense. [7] The company says the model was trained for defensive cyber work and can autonomously find, validate, and patch critical software vulnerabilities. [1]

Argon is scheduled to launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at a 95% discount. [7] Google told Axios that the model is state of the art on certain coding and knowledge-work benchmarks and outperforms OpenAI’s GPT-6 Astra on several of them. [4]

Google says thousands of employees are already using Argon for specialized coding tasks, deeper research, and writing, with the model powering internal workflows. [7] In one quantum-computing example, Argon helped researchers optimize the spacetime resources of bottleneck subroutines and beat a published baseline by 40% within minutes. [7]

A team of Argon agents analyzed fleet-wide profiling telemetry and applied memory optimizations that freed more than 300 TiB across Google’s data centers after deployment. [7] Google estimates those memory changes could ultimately save between 500 TiB and 1 PiB, while CNBC reported that the work avoided the need to purchase additional hardware. [7][10]

Across 18 benchmarks disclosed by Google, VentureBeat found that Argon led outright on 12 and tied for first on one. [9] In the same comparison, GPT-6 Astra led outright on three benchmarks and tied Argon on one, while Claude Opus 5.5 led outright on two. [9]

Axios cited Bloomberg’s report that some Google employees found Argon lacking in internal testing, while noting that Google called the characterization inaccurate. [4] The launch comes nearly a year after Gemini 3, during a period when Google continued releasing smaller and cheaper Flash models as OpenAI and Anthropic advanced their flagship offerings. [4]

Argon’s debut separates model announcement from general availability: Google is placing its strongest cybersecurity claims in the hands of selected defenders while it develops guardrails and participates in government review. Its disclosed comparison favors Argon overall, but GPT-6 Astra and Claude Opus 5.5 still lead some benchmarks, making Google’s advantage broad rather than universal. The internal memory and quantum examples provide operational evidence beyond benchmark scores, although the disputed employee feedback makes independent production testing especially important. What the sources don't address: when broader access will begin or whether independent real-world evaluations will reproduce Google’s reported advantages.