On September 30, Google introduced its strongest artificial intelligence model to date — Gemini 4 Argon. The company said the model is aimed at a broad range of tasks — coding, research, and text generation — but its main specialization is cybersecurity.

The most notable aspect of the announcement is that the model is not open to the public. Google is handing Argon only to selected cybersecurity partners through the Fairwind program. The reason is stated openly: capabilities at this level in the wrong hands could be used for cyberattacks on banks, hospitals, and government systems.

What was announced

Gemini 4 is Google's new flagship model family in the Gemini generation, and Argon is its strongest version. In a blog post by Google DeepMind senior vice president and the company's chief AI architect Koray Kavukcuoglu, the model is said to be designed to "maintain deep reasoning in complex, long-horizon workflows."

Three main directions of the model are highlighted:

  • Software engineering — writing code, debugging, and migrating large codebases;
  • Cybersecurity — independently finding, verifying, and fixing vulnerabilities;
  • Multimodal analysis — analyzing the content of long videos and diagrams.

Google's own employees are already testing the model in daily work: it's reportedly being used in debugging and codebase migration.

Why access is limited

Argon is given only to selected cybersecurity partners through Google's Fairwind program — the company's security initiative. The model was specifically trained for defensive cyber work, and the company claims it can "independently find, verify, and fix critical software vulnerabilities."

"Safely releasing frontier capabilities at this level requires a staged approach." — Koray Kavukcuoglu, Google DeepMind senior vice president and chief AI architect

Google is voluntarily giving the U.S. government early access to the model and gathering tester feedback before wider distribution. This approach mirrors rival Anthropic's path: the company keeps its most advanced model, Claude Mythos Preview, open only to a small group of trusted organizations. In June, Washington forced Anthropic to temporarily halt access to the publicly released Claude Mythos and Claude Fable models; since then a voluntary review process has been in place before releasing the strongest AI models.

The announcement came a day after U.S. President Donald Trump hosted tech leaders at the White House, including Google chief Sundar Pichai and Anthropic chief Dario Amodei, and signed a voluntary agreement in which they pledged to control the risks of their AI systems themselves.

Cybersecurity experts fear the most advanced technologies will be used to breach banks, hospitals, and government systems. Google emphasizes that Argon is designed to refuse requests that help carry out cyberattacks or create chemical, biological, and nuclear weapons — OpenAI and Anthropic also built such safeguards into their most advanced models.

Benchmarks and competition

Google claims in its blog that Argon showed significantly higher results than OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models across various AI benchmarks. Citing the increasingly popular benchmark startup Vals, the company reports that Argon now leads its AI models index.

These claims should be treated with caution: every lab picks the benchmark convenient to itself and puts itself first. Without independent verification, such results remain at the level of the company's own statements.

Meanwhile, the competitive environment is heating up. Recently OpenAI released Astra, called its best model, and Anthropic introduced Fable earlier this year — in both cases the rhetoric is nearly identical. Google, once considered "behind" in the AI race (NPR's assessment), has recently been succeeding with Gemini: in August the company announced the Gemini app has over a billion monthly users — making it competitive with OpenAI, which recently reported a billion monthly ChatGPT users.

What it means for cyber defense

Argon's strongest practical proof is early testers' experience. Per Google, testers used the model to find a vulnerability in software used in hospitals worldwide: the vulnerability exposed confidential personal data, and other advanced models couldn't find it.

This is a concrete example of AI working on the defense side: the model is deployed not as an attack weapon but as a tool in defenders' hands. Google said it will give Argon to trusted defenders and its internal teams without cyber restrictions — i.e., full use of the model's capabilities is allowed for defensive purposes.

At the same time, exactly this dual nature — a strong defense tool can also be a strong attack tool — is the main reason for the limited release. The industry is now moving from "announce first, think later" to "test first, then open."

Next steps: when it goes public

Google didn't name an exact date for wide distribution. The plan: first gather evaluations from trusted cyber defenders and testers, harden safety mechanisms, participate in the U.S. government's voluntary pre-release early-access process — and only then will the turn come for paid API clients and Google AI Ultra subscribers, then developers, businesses, and ordinary consumers.

This is a notable departure from the familiar scheme where a flagship model announcement came with a broad web or API release. The announcement is now not a product presentation but a statement about technical direction and access policy.

The Uzbek context

This news has several practical dimensions for Uzbekistan.

First, cybersecurity. Digitalization is accelerating in the country: banks, government services, and medical systems increasingly rely on digital infrastructure. If models like Argon become a strong tool in defenders' hands, Uzbekistan's banks and government systems will sooner or later use such defense tools too — but at the same time attackers also aspire to these technologies. For local CERT teams and bank security departments, this is a signal to prepare.

Second, developers and startups. The number of teams in Uzbekistan building products on AI APIs is growing. Frontier models no longer launch on the "announced — tomorrow in API" scheme: limited, staged release is becoming the new norm. When planning a product, it's wiser to rely not on the newest model but on a model with stable access.

Third, policy example. In the U.S., a voluntary review process between government and companies is taking shape. As Uzbekistan develops its national AI strategy, such "safety-first" release models can be an example for local regulators too.