Google presents flagship model Gemini 4 Argon

Google has announced its flagship neural network model Gemini 4 Argon, capable of generating up to 1 million output tokens in a single run. The new model is designed to solve long, multi-stage tasks in software development, cybersecurity, as well as financial and legal analysis. The previous output limit was 64,000 tokens. According to Google, the expanded volume allows AI agents to perform complex sequences of actions without breaking down the work into separate requests.

Within the corporation, Gemini 4 Argon is already being used to optimize data centers, where the proposed solutions have allowed the release of over 300 TB of RAM with the potential to save up to 1 PB. In addition, the model's agents have been involved in optimizing quantum algorithms and translating system code from C/C++ to Rust, including over 800,000 lines of the Zircon kernel of the Fuchsia operating system and 32,000 lines of SIMD code of the libgav1 video decoder.

The wide launch of the product has been postponed: first, the system will be deployed for trusted cybersecurity specialists as part of the Fairwind program. Later, access will be opened to Gemini API users and Google AI Ultra subscribers.