Google launches a new generation of cutting-edge model Gemini 4 Argon to enhance long-term task processing and enterprise automation capabilities

Google today (1st) announced the launch of a new generation of cutting-edge model Gemini 4 Argon, which is designed to handle complex, long-term and multi-step workflows. Its applications include enterprise knowledge work such as software engineering, finance and law, and network security defense. Currently, Google is opening it to trusted network security defense experts for testing through the Fairwind project, and will gradually expand to developers, enterprises, and consumers in the future.

A major upgrade of Gemini 4 Argon is to greatly improve the model's ability to handle long tasks. Google has significantly increased the upper limit of output tokens from the original 64,000 to 1 million, allowing the model to perform longer reasoning in a single task and complete more complex multi-step work.

Argon is now integrated into Google's internal workflow. Google stated that thousands of employees use Argon in tasks such as professional program development, in-depth research, and writing. In terms of data center memory optimization, the Argon agent can independently analyze system performance and apply improvement plans. After full deployment, more than 300 TiB of memory has been released, and it is estimated that the overall memory capacity can be saved about 500 TiB to 1 PiB.

The Argon agent is also used to migrate Google's internal C/C++ code library to Rust, covering core libraries with tens of thousands of lines of code, and even the Fuchsia Zircon core with more than 800,000 lines of code. Taking the video decoding open source software libgav1 as an example, after Argon processed about 32,000 lines of SIMD code, the execution speed of the new version was 2.7 times faster than the original Rust version, and was close to the highly optimized C++ version.

In addition to program development, Google also targets professional jobs such as finance, law, taxation, and enterprise automation with Argon. In the DeepSWE v1.1 test, which evaluates real-world long-term software engineering tasks, Argon achieved 77.9%; Zapier's AutomationBench achieved 51.3%. In the long video understanding LVBench test, the score reached 91.7%.

Network security is another major capability of Argon. Google said that Argon can independently find, verify and repair software vulnerabilities. The security company Wiz has put it into testing through the "Scan for Good" project. Argon once identified a critical vulnerability that could lead to the leakage of sensitive personal data in medical software; in the CWE-bench v1 test that evaluates vulnerability patching capabilities, it tied for the highest score of 68%.

However, as AI agents can perform longer and more complex autonomous tasks, Google has also strengthened security measures, including preventing the model from being maliciously exploited, resisting prompt injection attacks, monitoring whether the model deviates from the user's original intention, and strengthening the sandbox environment used when the AI ​​agent is executed.

In terms of price, Gemini 4 Argon charges US$2 per 1 million input tokens (word elements) and US$10 per 1 million output tokens. The cache input token price is 95% lower than that of general input. Google said that Argon is currently open in phases and will be available to paying API customers and Google AI Ultra subscribers first in the future.