Google launches Gemini 4 Argon with focus on cybersecurity

Google is putting cybersecurity at the heart of its latest Gemini release, giving a limited group of security partners access to Argon. The model is designed to identify and repair software vulnerabilities while also handling coding, research and visual tasks, as the company steps up competition with OpenAI and Anthropic.

Google has introduced Gemini 4 Argon, a new AI model that the company says is designed to tackle demanding coding, research and writing tasks, while placing particular emphasis on cybersecurity.

The model is not being released broadly at this stage. Google is making Argon available to a limited set of cybersecurity partners through its Fairwind Program, allowing selected organisations to test its capabilities in defensive security work.

According to Google, Argon was trained with cybersecurity applications in mind and can identify software weaknesses, check whether those vulnerabilities are genuine and develop fixes with limited human intervention. The company describes the model as capable of "autonomously find, validate, and patch critical software vulnerabilities".

The security focus comes as AI companies increasingly explore whether their latest systems can move beyond assisting security researchers to handling parts of the vulnerability discovery and remediation process themselves. Such capabilities could potentially speed up defensive work, although Google's claims are based on its own testing and early use of the model.

Google puts coding and long-running tasks at the centre

Cybersecurity is not the only area where Google expects Argon to be used. The company says employees have already been applying the model to everyday engineering tasks, including debugging and moving software between codebases.

Google has also highlighted Argon's ability to work with visual information. That includes interpreting charts and examining material contained within lengthy videos, alongside its text and coding capabilities.

The company positions the model as one intended to maintain reasoning over complicated tasks that take time to complete, rather than simply answering individual prompts. In a blog post announcing the system, Google said Argon was "fundamentally changing the way we work and build at Google".

Those claims will be closely watched as the leading AI developers continue to compete on increasingly broad measures of model performance. Google said its internal and external evaluations showed Argon outperforming OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models across several benchmarks.

Google cited Vals, an AI benchmarking company, as evidence for that comparison, saying Argon currently sits at the top of its model index. Benchmark results, however, can vary depending on the tests, scoring methods and versions of the competing systems involved.

Gemini's momentum adds to Google's AI push

Argon's arrival comes during a period in which Google has sought to close the gap with rivals in generative AI. The company has increasingly integrated Gemini across its products while continuing to develop models aimed at specialised professional uses.

Google said in August that its Gemini application had surpassed one billion monthly users. OpenAI has reported a similar milestone for ChatGPT, underlining the scale of the competition between the two companies.

At the same time, leading AI developers have continued releasing more capable systems while publicly discussing the risks associated with increasingly powerful AI. OpenAI recently promoted Astra as its most capable model, while Anthropic has also introduced newer models aimed at expanding performance across complex tasks.

For now, Argon's limited availability means Google's claims will largely be tested through the partners receiving access rather than through broad public use. Its cybersecurity capabilities, in particular, are likely to attract attention as companies assess whether AI can reliably discover and repair real-world vulnerabilities without creating additional security risks.