Google launches Gemini 4 Carbon for cybersecurity defense

A new tool for defense

Google is introducing Gemini 4 Carbon, a model specifically designed to assist cybersecurity defenders in identifying and fixing software flaws. Unlike models that only generate text, Carbon is built to handle engineering tasks that require sustained work across multiple steps. This includes tracing bugs through large, interconnected codebases to find the root cause of a problem.

The model is currently being released to a select group of cybersecurity defenders for early access. According to EdTech Innovation Hub, this rollout allows experts to test how the AI handles complex software vulnerabilities before it reaches a wider audience. Google plans to expand access through the Fairwind Program before releasing it to the general public.

The goal of Carbon is to help security teams distinguish between a minor weakness and a flaw that could actually be used to compromise a system. By providing verified patches, the model aims to help engineers prioritize which fixes are most urgent. This focus on defensive security is a primary differentiator for the new release.

Performance and benchmarks

Google reports that Carbon achieved a score of 59 on version 5.0 of the Artificial Analysis Intelligence Index. This debut position highlights the model's improvements in reasoning and its ability to perform complex scientific and mathematical tasks. Beyond security, it also reached fourth place on the WebDev Arena at launch.

In specific security testing, Carbon scored 88.2 percent on SEC-bench Pro, which evaluates the ability to uncover vulnerabilities in real software. It also scored 54.8 percent on ExploitGym, a benchmark that assesses whether a discovered vulnerability can actually be exploited. These scores were achieved without using guardrails, which are the restrictions typically used to control a model's behavior.

The model also showed significant progress in repair capabilities, reaching 71 percent on Collinear’s CWE-bench. Internally, Google found that Carbon more than doubled the success rate of the previous Argon model on tests for generating secure code. The company noted that engineering teams have used the model to work on difficult tasks for days, investigating dead ends and tracing bugs across large projects.

Safety and reliability

Google is focusing on making Carbon more reliable by improving how it follows instructions and handles uncertainty. Internal testing showed that Carbon produced roughly half as many unsupported claims as the Argon model. Furthermore, the frequency of serious alignment warnings dropped by nearly 80 percent when used in the live coding environments of Google's own staff.

The model also performed well against specific types of cyberattacks, such as indirect prompt injections. These attacks occur when an AI system is tricked into following malicious instructions hidden in documents or web pages. Carbon recorded a zero percent attack success rate on the Gray Swan benchmark designed to test this specific vulnerability.

To ensure the model remains safe for public use, Google is gathering feedback during this initial rollout. The company is also participating in a voluntary process by the US government to test models before they are released to the public. These measures are intended to refine the model's safeguards as it moves toward a broader release for businesses and consumers.

Pricing and rollout plan

The pricing for Gemini 4 Carbon will initially match the introductory rates of Gemini 4 Argon. This means it will cost $2 for every million input tokens and $10 for every million output tokens. Tokens are the basic units of information that a model processes, where input covers what it receives and output covers what it generates.

Users who reuse previously processed material can take advantage of a 95 percent discount on input rates for cached tokens. However, Google noted that these rates will eventually increase to $4 per million input tokens and $20 per million output tokens once the introductory period ends. Clément Farabet noted that Carbon delivers stronger reasoning without increasing the token pricing over Argon.

The rollout is structured to prioritize organizations that defend critical systems. The Fairwind Program will provide early access to governments, healthcare providers, and telecommunications services. Academic laboratories focused on defensive benchmarking are also being invited to participate in the program to help refine the model's capabilities.