Gemini 4 Carbon (blog.google)

Today marks the launch of Gemini 4 Carbon, the latest frontier AI model, boasting enhanced cybersecurity and robust long-horizon capabilities. With a score of 59 on the updated Artificial Analysis Intelligence Index v5.0, Carbon excels in coding and cybersecurity tasks, achieving a remarkable 88.2% on SEC-bench Pro for vulnerability detection and over 2x improvement in generating secure code compared to its predecessor, Argon. This model is designed to tackle complex and multi-step engineering challenges, thereby streamlining the software development lifecycle and improving output quality across various domains, including full-stack web development. Gemini 4 Carbon also prioritizes safety and security enhancements, achieving 0.0% successful attempts at prompt injection attacks and reducing high-severity misalignment alerts by nearly 80%. The model adopts a user-centric approach, facilitating better collaboration and transparency while mitigating risks of shortcuts and unsupported claims. The introduction of Carbon represents a significant step forward for the AI/ML community, as it combines advanced capabilities in software engineering and cybersecurity with a strong commitment to alignment and safety standards. Access to the model will be gradually rolled out to select cyber defenders and trusted partners through a structured Fairwind Program, ensuring thorough testing and feedback before broader release.