Google DeepMind launches Gemini 4 Argon, its new generation for cyber-defense partners with a one-million-token output window

Google DeepMind announced on September 30, 2026 its new flagship model Gemini 4 Argon, marking the first transition to a new generation of its core models since Gemini 3 was launched last November. Korai Kafoktsiuoglu, who runs daily operations at Google DeepMind, described the model as the next era of advanced intelligence at the company, designed to handle long and complex workflows across software engineering, knowledge work in the legal and financial sectors, as well as defense in cybersecurity.

Google began making the model available to a limited group of cyber defense partners and U.S. government agencies through the Virwind program, and decided to provide these defenders with the model without the usual cybersecurity safeguards.so they could fully leverage its technical capabilities in vulnerability detection. Cybersecurity firm Waze recorded early use of the model in its Scan for Good initiative, where Argon discovered a critical vulnerability in medical software used by hospitals worldwide, a flaw that previous leading models failed to detect.

In terms of benchmark metrics, Argon achieved 77.9 percent on the Deep S SWE 1.1 score, which focuses on extended software engineering tasks, and shared first place with 68 percent on the C SWE Bench metric that measures models' ability to remediate security vulnerabilities. Google backed these capabilities with an architectural leap that raised the maximum output from 64,000 tokens to a full million tokens, while the company already relies on the model internally, with thousands of employees using it, and independent Argon agent teams have offloaded more than 300 terabytes of memory capacity through Google data centers.

The company set a preliminary price of $2 per million input tokens and $10 per million output tokens, announcing that paid API customers and Google AI Ultra subscribers will be the next group to gain access. Sander Betchai confirmed that the model includes advanced security guarantees and will soon expand to developers and users, although the company has not yet announced an official date for opening general access to developers or consumers.

This technological shift has direct implications for chief technology officers and software security teams in the region, particularly in the Gulf markets, Egypt, and the Greater Arab area. Raising the output ceiling to one million tokens, together with the preliminary inference cost of ten dollars per output million, practically means that financial institutions and banking system developers in Riyadh, Dubai, and Cairo can abandon complex code fragmentation and rely on a single-step audit of entire code repositories. Moreover, the model's unrestricted access to cyber defense teams confirms that the next technical battle over regional infrastructure will require local engineers to upgrade their automated code review skills, in anticipation of an automated wave of software vulnerability discoveries before complex testing tools can catch up.