Google DeepMind Unveils Gemini 4 Argon with 1M Token Output Limit and Cybersecurity Focus

A major technical update in Argon is the significant expansion of its output token limit to 1 million tokens, up from the 64,000 tokens available in previous Gemini models. This increased headroom is intended to allow the model to think deeply and generate extensive responses in a single trajectory to solve complex problems.

In performance evaluations, Argon has demonstrated high scores across several benchmarks. It achieved 77.9% on DeepSWE v1.1, a benchmark for real-world long-horizon software engineering tasks, and ranks #1 on AutomationBench for end-to-end execution across core business functions with a score of 51.3%. The model also achieved a 91.7% score on the LVBench benchmark for long video understanding. Additionally, Argon tied for first place on CWE-bench v1, which evaluates the ability to remediate security vulnerabilities, with a score of 68%.

To bolster cybersecurity capabilities, Argon is trained to autonomously find, validate, and patch critical software vulnerabilities. Google is currently rolling out the model to a limited group of trusted defenders through its Fairwind Program and is participating in the U.S. government’s voluntary process for pre-release model access. In an early demonstration, the security company Wiz used Argon to uncover a critical vulnerability in healthcare software used by hospitals worldwide, which Google claims previous frontier models had missed.