Google DeepMind Announces Gemini 4 Carbon Frontier Model with Cyber-Security Focus

Imagine your GP surgery, your local council, or the systems that keep the lights on suddenly finding a critical software flaw before a criminal does. That’s the kind of work Google DeepMind says its new Gemini 4 Carbon model is built for. Announced in October 2026, Carbon is Google’s latest frontier AI model — and the company is pitching it squarely at the world of cyber defence.

It’s a bold claim, and one that deserves a close look.

What Is Gemini 4 Carbon?

Carbon is the second model in Google’s Gemini 4 family, arriving roughly a fortnight after Gemini 4 Argon. Google DeepMind describes it as its strongest model, with a particular focus on cyber security and what the company calls “long-horizon” tasks — meaning complex jobs that require sustained planning and the completion of many dependent steps in sequence, rather than a simple one-shot answer.

Think of it less like a search engine and more like a very capable assistant that can work through a multi-stage engineering problem over an extended period without losing the thread.

Google says Carbon is designed for agentic workflows — situations where an AI system takes a series of actions autonomously, checking its own progress and adapting as it goes. In coding and software engineering, that means it can potentially write, test, debug, and patch code with far less human hand-holding than earlier models required.

The Benchmark Numbers

Google has published three benchmark scores to back up its claims. On SEC-bench Pro, a test for finding software vulnerabilities, Carbon scored 88.2%. On ExploitGym, which tests exploitation capabilities, it scored 54.8%. And on CWE-bench, which measures how well a model can patch known weaknesses, it scored 71.0%.

Those are striking figures — if they hold up.

It’s worth being clear here: these benchmarks were reported by Google itself. Independent researchers have not yet confirmed the methodology, the test conditions, or how Carbon compares with rival models from Anthropic, OpenAI, or Meta under the same conditions. The “state-of-the-art” label is Google’s own assessment. That doesn’t mean it’s wrong — but it does mean the numbers deserve scrutiny before anyone treats them as settled fact.

Who Can Actually Use It?

Not everyone, at least not yet. Google is rolling out Carbon through controlled access, initially limited to a subset of trusted cyber-security defenders. Wider availability is planned through Google’s Fairwind programme, though the company hasn’t published a firm timeline for unrestricted access.

That cautious approach makes sense given what the model is capable of. A system that scores 54.8% on an exploitation benchmark could, in the wrong hands, be pointed at vulnerabilities rather than used to fix them. Controlled release gives Google and its partners time to assess real-world behaviour before opening the floodgates.

What Does It Cost?

Google has published pricing in US dollars: an introductory rate of around $2 per million input tokens and $10 per million output tokens (roughly £1.60 and £8 respectively at current rates, though Google has not confirmed official sterling pricing). The Register reports that the rates rise to $4 and $20 after the introductory period. Cached input tokens — previously processed data that the model can reuse — are reported to receive a 95% discount on the standard input rate, which could make sustained, repeat-use tasks considerably cheaper for organisations running large workloads.

But pricing is only part of the picture. Any organisation handling sensitive data — chiefly in healthcare, local government, or critical infrastructure — would need to satisfy itself that Carbon meets data-protection, procurement, and security requirements before deployment. That process takes time and doesn’t happen automatically just because a model becomes commercially available.

Cyber Security: Opportunity and Risk

The dual-use nature of a model like Carbon is impossible to ignore. Better tools for finding and patching vulnerabilities are genuinely useful — for banks, hospitals, utilities, and government departments. But a system capable of identifying and validating exploits is also a system that, misused or poorly secured, could assist attackers.

Google’s decision to limit initial access to defenders is one response to that tension. Whether it’s sufficient is a question that independent security researchers and regulators will likely press in the months ahead.

And there’s a broader point. More capable AI systems don’t automatically make organisations safer. They require skilled people to interpret outputs, governance frameworks to manage risk, and procurement processes that actually work. The technology is only one piece.

What This Means for Kent Residents

There’s no confirmed deployment of Gemini 4 Carbon by any Kent organisation at this stage, and no direct effect on local jobs, services, or household costs has been identified. But bodies like Kent County Council, Medway Council, and NHS Kent and Medway ICB all manage large volumes of sensitive data and face real cyber-security pressures — so tools of this kind will eventually be relevant to them, subject to the appropriate governance and data-protection checks. For now, Carbon is one to watch rather than one to act on.