Google Unveils Gemini 4 Carbon, Starting with Security Professionals
The new model is focused on long-running tasks and cyber defense. Access is limited at first to a select group of security professionals.

Google DeepMind has introduced a new AI model, Gemini 4 Carbon. The company says it was designed for agentic work, meaning tasks that an AI carries out on its own across multiple steps, as well as for cyber defense. For now, only selected security professionals can use it.
Carbon arrives roughly two weeks after Gemini 4 Argon. The Decoder reported that it took first place on the Artificial Analysis Intelligence Index v5, a benchmark that bundles several capabilities into a single score, with 59 points. Ars Technica, by contrast, described it as a modest step up from Argon.
Google had previously showcased Argon organizing enormous volumes of material and generating 3D content. It turned more than 80,000 NASA files accumulated over 72 Mars flights into a presentation, and built a playable 3D track from photos of a real desk. Carbon adds cyber defense capabilities on top of that same ability to sustain complex work over long stretches.
Source: Google (press release)
For everyday users, the biggest difference between the two models is who can actually use them. Argon is available to developers and enterprises, as well as to Google AI Ultra and Pro subscribers. Carbon is landing in the hands of security professionals first, so the questions to watch next are what it delivers in real security work and when access will widen.
