Google Begins Limited Rollout of Gemini 4 Carbon
Google DeepMind is giving selected cyber defenders early access to Gemini 4 Carbon, a new model it describes as its strongest system yet for cybersecurity and long-horizon engineering.
By RobotPulse Newsroom ·
Watch the RobotPulse Short on YouTube
Google DeepMind has announced Gemini 4 Carbon, a new model aimed at long-running engineering work and cybersecurity defense. Rather than opening it to everyone at launch, Google says it is starting with a limited group of cyber defenders and will expand access gradually.
What happened
Google DeepMind describes Carbon as its strongest model so far for cybersecurity and long-horizon capabilities. The company says Carbon leads the updated Artificial Analysis Intelligence Index v5.0 with a score of 59, and reports high results on several security benchmarks that measure vulnerability discovery, validation and remediation. Those are company-reported results, and they should be read as benchmark claims rather than a guarantee of performance in every real-world setting. Google DeepMind
The rollout is deliberately staged. Google says a subset of cyber defenders will receive access first, with additional trusted partners added through its Fairwind Program. Wider access for developers, businesses and consumers is planned for later, after the company gathers feedback and continues work on its safeguards. Reuters also reported that the new model would initially be available to cybersecurity defenders. Reuters
DeepMind says Carbon is built for sustained, multi-step work, including engineering tasks that may require extended reasoning across large codebases. The company has also set introductory API prices at $2 per million input tokens and $10 per million output tokens, matching the introductory pricing it lists for Gemini 4 Argon. Its announcement does not mean that general access is open today.
Why it matters
The announcement links two increasingly important areas of model development: agents that can work through complex tasks over time, and models that can assist with both finding and fixing software weaknesses. Google is presenting the defensive uses as central to Carbon’s purpose, while its own benchmark descriptions also measure capabilities that could be sensitive if used outside authorized security work.
A limited release gives Google a way to gather feedback before broad deployment, but it also makes independent testing and rollout details important. The model’s benchmark position is one signal, not a complete measure of safety, reliability or value to defenders. Real-world usefulness will depend on how it performs in authorized security workflows, whether its safeguards hold up under pressure, and how it handles tasks beyond the benchmarks highlighted by the company.
What to watch next
The near-term questions are who qualifies for the initial access group, how quickly trusted partners are added, and when broader developer access begins. Independent assessments of Carbon’s benchmark results and security behavior will also help separate the company’s reported gains from performance that outside researchers can reproduce.
Google says it will use feedback from early users to iterate on safeguards before bringing the model to its wider product surfaces. That staged path is the main development to follow: the announcement introduces a new frontier model, but the wider impact will depend on the evidence and access decisions that come next.