Google's move to shake up the AI world: Gemini 4 Argon unveiled

Google has unveiled its new AI model, Gemini 4 Argon. The company will proceed with a limited test group rather than opening the model to a wide user base in the initial phase. Initial access will be granted to specific cybersecurity experts; subsequently, a phased plan will be followed for access by developers, companies, and individual users.

Gemini 4 Argon is positioned specifically for use in long and complex tasks. According to Google's statements, the model was developed to excel in areas requiring intensive information processing, such as software development, finance, law, research, and cybersecurity.

The company has also begun using Argon in its own internal processes. It was stated that Google employees are evaluating the model for code development, research, and engineering work, while an example regarding memory usage in data centers drew attention. It was reported that Argon-based systems enabled the reuse of more than 300 TiB of memory, with the total potential savings calculated to be between 500 TiB and 1 PiB.

Google is also utilizing the model for large code projects. The company is reportedly using Argon in its efforts to migrate some C and C++ code to the Rust language, a scope that includes more than 800 thousand lines of code in the Zircon kernel of the Fuchsia operating system.

One of the most notable features of Gemini 4 Argon is the significant increase in its output limit. The output limit, which remained at 64 thousand tokens in previous Gemini models, has been increased to 1 million tokens in Argon. Google states that this allows the model to complete long tasks without breaking them into pieces.

This capacity is particularly important for tasks such as multi-stage coding processes, comprehensive research, and the examination of large document groups. According to test results shared by the company, Gemini 4 Argon scored 77.9 percent on the DeepSWE v1.1 test, which measures real-world software development tasks. Google also states that the model provides high results in fields such as finance, law, automation, and long-form video analysis.

One of the first areas of use for the model will be cybersecurity. Google is opening Gemini 4 Argon to a limited number of experts in the Fairwind program. Cloud security company Wiz was also mentioned as being among the organizations testing the model.

According to Google's statement, Argon detected a serious security vulnerability in software used in hospitals that could have led to the exposure of personal information. The company claims that previous advanced models failed to find the same vulnerability.

Google also announced that it is conducting additional security tests before moving to general use due to the model's powerful cybersecurity capabilities. It was noted that different protection systems have been developed against risks such as misuse, command injection, and the model deviating from user requests.

It has not yet been finalized when Gemini 4 Argon will be opened for general use. Google will first collect feedback from the test group. In the next phase, paid API users and Google AI Ultra subscribers will be able to access the model.

According to the initial pricing, a fee of 2 dollars will be charged for 1 million input tokens and 10 dollars for 1 million output tokens. It was announced that after the introductory period, these prices will increase to 4 dollars and 20 dollars, respectively.