Google DeepMind Releases 'Gemini 4 Carbon'

Google DeepMind has released a new artificial intelligence model 'Gemini 4 Carbon.' The model recorded a score of 59 on the artificial intelligence performance evaluation metric 'Artificial Analysis Intelligence Index v5,' currently achieving the top rating. Its release came just two weeks after the previous model 'Gemini 4 Argon.'

The momentum of the Gemini 4 series has accelerated over the past few weeks. The brief interval of two weeks between Gemini 4 Argon's release and Carbon's public unveiling demonstrates that Google DeepMind is significantly compressing its development cycle. The company's proactive stance in AI model performance competition is evident from these consecutive releases.

Google DeepMind has identified two primary use cases as design features of Gemini 4 Carbon. One is support for 'extended agent workflows,' with design optimized for extended operations where AI autonomously processes multiple stages. The other is cybersecurity applications, intended to cover the entire defensive lifecycle.

Due to its advanced security capabilities, it remains limited to selected security experts rather than general public release at this stage. Given the potential for AI models to be leveraged in both cyber attacks and defenses, a phased access management approach appears necessary to prevent inappropriate use.

The application of AI to cybersecurity is a field gaining increased industry attention, and Gemini 4 Carbon is positioned as a product aligned with this trend. Notably, as the phrase 'entire defensive lifecycle' suggests, the concept of AI consistently supporting threat detection through response and recovery represents a potential new competitive axis in the enterprise security market.

Meanwhile, the top evaluation on the Artificial Analysis Intelligence Index also represents an indicator of high general-purpose performance. While possessing foundational model capabilities not limited to security applications, the approach of initially narrowing distribution to experts can be understood as managing risks while accumulating knowledge.

Future points of interest include how the limited distribution's scope will expand and the timing of when access for general users and enterprises will be opened. As evaluations from security experts accumulate, the actual effectiveness and challenges of the model are expected to become clearer. Additionally, whether the two-week release interval continues or what systems of models will subsequently emerge are points the industry should monitor.