Google Unveils Powerful AI Model Gemini 4 Argon with Israeli Cybersecurity Collaboration

Google DeepMind has announced its latest flagship artificial intelligence model, Gemini 4 Argon, the first in its fourth generation. This new model boasts an unprecedented ability to generate a continuous output of one million tokens, a significant leap from previous models limited to 64,000 tokens. This enhanced capability allows AI agents to perform extended research, write entire code libraries, and solve complex organizational problems in a single, unbroken sequence.

Gemini 4 Argon has demonstrated superior performance in various benchmarks, outperforming OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 in areas like autonomous software engineering and financial/tax/code tasks. However, it still trails in some command-line and operating system tasks. Internally, Google engineers are already using Argon to optimize quantum calculations, save significant server memory, and convert code for its Fuchsia operating system.

A key aspect of the launch is a close collaboration with Israeli cybersecurity firm Wiz. Through the Scan for Good initiative, Wiz researchers utilized Argon to identify critical security vulnerabilities in medical software used by hospitals worldwide, flaws that previous models had missed. In blind penetration tests, Argon achieved a 70.9% success rate in attacking live websites without source code access.

Google is adopting a cautious rollout strategy for Gemini 4 Argon, initially making it available only to a select group of cybersecurity partners under the Fairwind program. This measured approach is due to the model's exceptional power in offensive cyber capabilities and engineering, which necessitates careful deployment to prevent misuse. The company is also undergoing voluntary review with U.S. government and security agencies before wider release.

Google highlights Gemini 4 Argon's robust security features, including exceptional resistance to indirect prompt injection attacks, with only a 0.7% breach rate in rigorous testing. The model also incorporates enhanced monitoring of its thought process to prevent agents from deviating from their tasks. Pricing for the model will be competitive, with an introductory rate of $2 per million input tokens and $10 per million output tokens, with significant discounts for cached reads. Future access will be expanded to API developers and Google AI Ultra subscribers before integration into consumer products.