Google introduced Gemini 4 Argon. The most powerful model has so far only been given to white-hat hackers

At the end of September, Anthropic and OpenAI released new models, and now Google’s response has arrived. Gemini 4 Argon opens a new generation of Gemini and, according to Google’s Head of AI Architecture Koray Kavukcuoglu, changes the way Google works. But there’s a catch: you can’t try it yet. Google has entrusted it only to a small group of cybersecurity experts.

First defenders, then others

According to Google, Argon can independently find, verify, and fix critical security vulnerabilities in software. The same capability in the wrong hands would serve attackers, which is why Google is opting for gradual access. The first to receive the model were members of the Fairwind program, which Google launched on September 2 for governments, critical infrastructure operators, and technology companies. According to available information, it has over 650 partners. These defenders and Google’s internal teams will receive Argon without cyber safeguards to fully utilize its capabilities. Competitors have previously chosen a similar path, for example, Anthropic with the Claude Mythos Preview model as part of Project Glasswing.

The first practical result was described by the security firm Wiz. According to them, Argon uncovered a critical vulnerability in hospital software used worldwide, which threatened the leakage of sensitive personal data. Previous models overlooked it. Google did not disclose which software it was. For others, Argon will come later, first for paying API customers and Google AI Ultra subscribers. A date is not yet available.

A million tokens in one go

The most significant technical change is the output limit. Previous Gemini models could generate a maximum of 64 thousand tokens in a single response, while Argon can handle a million. The model’s “thinking” is also counted towards the limit, so it’s not about longer responses, but about more space to ponder a complex task. Argon can thus process extensive analysis or a large code modification at once, instead of dividing it into dozens of sequential steps.

Google is already building on it

Inside Google, thousands of employees are using Argon, according to the company. Agents built on the model are converting old code from C and C++ languages to safer Rust, including over 800 thousand lines of the Fuchsia system kernel. For the libgav1 video decoder, they created a Rust version that is 2.7× faster than the previous attempt. Another team of agents analyzed data center operations, and their optimizations freed up over 300 TiB of memory. However, all these are Google’s figures and have not been independently verified.

Leads in tables, but not everywhere

Google boasts first place in the DeepSWE v1.1 test, which measures the handling of long programming tasks, and in the AutomationBench business process test with a result of 51.3%. However, Google’s own comparison also shows losses:

Moreover, what we already wrote about the competition applies: each company measures differently. For example, Google states a value of 57.9% for the Claude Fable 5.1 model in Terminal-Bench 4.0, while Anthropic’s own tables show 55.8%. Therefore, an unambiguous winner will only be determined by independent tests.

Introductory price like GPT-6 Sol

During the introductory period, developers will pay 2 dollars per million input tokens and 10 dollars per million output tokens; repeatedly used input will be 95% cheaper. After the promotion ends, the duration of which Google did not specify, the price will double to 4 and 20 dollars. Argon will thus start at the level of the recently released GPT-6 Sol and end at the level of Claude Opus 5.5.

Monitoring thoughts

In addition to performance, Google emphasizes security. Argon is designed to refuse assistance with attacks and weapons of mass destruction. Google monitors its internal activities to detect misuse and continuously checks the model’s thought process and its steps to ensure it doesn’t go beyond the given task. An interesting detail: Google deliberately does not use findings from this oversight during training so that the model does not learn to hide its reasoning from inspection. The company has also joined a voluntary U.S. government program that allows it to test the model before release.

For regular users of the Gemini application, including Czech ones, nothing changes yet. Google has not said when and in which tariffs Argon will become available to them.

Is it right to entrust the most powerful AI models only to selected experts first?

Source: Google, tbreak, The Hacker News, How2Shout, The Hack Academy