Google Unveils Gemini 4 Carbon, a New AI Model Built for Cybersecurity and Complex Coding

Google has introduced Gemini 4 Carbon, a new frontier artificial intelligence model built for cybersecurity research, software development and complex autonomous tasks. The launch, announced on Wednesday, adds a specialist system to Google’s Gemini 4 family just two weeks after the arrival of Gemini 4 Argon.

Carbon is aimed at work that takes sustained effort rather than a single quick answer. Google says it is designed for engineering tasks that can require work over multiple days, including investigating software vulnerabilities, developing and repairing code, and carrying out long-running autonomous assignments with limited supervision.

What Carbon is built to do

Where general-purpose chat models answer questions, Carbon is positioned as a working tool for security teams and developers. Its stated strengths are finding vulnerabilities in software, developing exploits in controlled test settings, repairing vulnerable code and generating secure code. Those are the same capabilities that make the model sensitive: the skills used to defend a system are closely related to the skills used to attack one.

That dual-use tension runs through the whole release. Google is presenting Carbon as a defender’s tool first, and the rollout reflects that choice. Rather than opening the model to everyone at once, it is starting with selected cybersecurity organisations, a group best placed to test it against real defensive work and to report problems before wider use.

The restricted rollout is the story as much as the model itself: a system built to find flaws is being handed first to the people paid to defend against them.

The benchmark numbers — and their limits

Google reports that Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. On specialist security evaluations reported by Google, the model scored 88.2 percent on SEC-bench Pro for vulnerability detection, 54.8 percent on ExploitGym for exploit development and 71 percent on CWE-bench for software vulnerability repair.

Those figures are Google’s own reported results, and they should be read that way. Benchmark scores describe performance on specific tests, not a guarantee of how a model will behave on a company’s own code, in its own systems, or against unfamiliar threats. Independent evaluations will be needed to establish how Carbon performs across different applications, and Google itself reports improvements in generating secure code and reducing unsupported claims that still need outside verification.

The numbers are still useful as a signal of intent. Vulnerability detection, exploit development and repair are three different skills, and Carbon is being measured on all three at once. That combination is what separates a specialist security model from a general coding assistant, and it is why security teams will be watching the first independent tests closely.

A deliberately limited rollout

Carbon will first become available to selected cybersecurity organisations, with access gradually expanding through Google’s Fairwind Program. Google says it is also participating in voluntary government testing before broader deployment, a step that acknowledges the risks of releasing a model capable of discovering and exploiting software vulnerabilities.

The cautious approach mirrors a wider shift in how frontier models are released. As AI agents gain the ability to modify software, investigate vulnerabilities and operate computer systems with limited supervision, the question is no longer only what a model can do, but who gets it first and under what controls. Carbon’s staged release is Google’s answer to that question, at least for now.

Developers are already seeing this shift elsewhere in the industry. GitHub Copilot Fusion routes coding prompts to the models it judges suitable for each task, while a Chinese AI model tested for cyberattacks in controlled settings shows how security evaluation itself is becoming a headline capability. Carbon sits in the same trend: AI that does not just write code, but probes, tests and repairs it.

What it costs to use

Google’s introductory API pricing for Carbon matches Gemini 4 Argon at 2 dollars per million input tokens and 10 dollars per million output tokens. That pricing puts it in line with its sibling model rather than positioning it as a premium-priced specialist, which may help security teams experiment with it once access widens.

Price, however, is only part of the cost calculation for enterprises. Businesses are buying AI faster than they are making it pay, and a security model adds its own overhead: human review of its findings, testing of its suggested repairs, and governance over autonomous actions on live systems. The organisations that benefit first are likely to be those that already have those controls in place.

• Carbon follows Gemini 4 Argon, launched two weeks earlier, in Google’s frontier model family.
• Its reported security scores cover detection, exploit development and repair — three distinct skills.
• Initial access is limited to selected cybersecurity organisations through the Fairwind Program.
• Introductory API pricing matches Argon: $2 per million input tokens and $10 per million output tokens.

Why this release matters now

The release intensifies competition with OpenAI and Anthropic, especially as enterprises adopt agents that can modify software and investigate vulnerabilities with limited supervision. It also arrives at a moment when cybersecurity threats are growing more sophisticated and when businesses are questioning whether heavy AI spending is producing matching returns.

For security teams, the practical question is whether a model like Carbon can shorten the gap between discovering a flaw and fixing it. Faster detection only helps if repair keeps pace, and automated repair only helps if a human checks the fix. Carbon’s reported repair score of 71 percent on CWE-bench suggests progress, but it also leaves a large share of cases where the fix still needs human work.

What to Watch Next

Watch for the first independent evaluations of Carbon, for how quickly access expands beyond the initial cybersecurity organisations, and for what voluntary government testing concludes before broader deployment. The benchmark figures reported by Google will be tested against real codebases, and the results of that testing will matter more than the launch numbers. Also watch whether enterprises pair the model with strict human review, or let autonomous agents act on its findings directly, because that choice will shape how much trust a system like Carbon earns in practice.