Gemini 4 Carbon: Google’s New AI Model for Cybersecurity

Google unveiled Gemini 4 Carbon on Wednesday, October 28, introducing a new frontier Artificial Intelligence model designed to strengthen Cybersecurity by identifying software vulnerabilities, developing fixes and generating more secure code.

The launch marks Google’s second major AI model release in a month, following Gemini 4 Argon. According to the company, Carbon delivers improvements in coding, agentic capabilities and cybersecurity, with a particular focus on completing complex engineering tasks that may require extended periods of work.

Google’s benchmark results indicate that Gemini 4 Carbon outperformed OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in almost every test the company highlighted. However, benchmark results provide only a partial picture of real-world AI performance, and the findings cited at launch were provided by Google.

The company is taking a cautious approach to the rollout, initially making Carbon available to a select group of cybersecurity defence organisations before expanding access through its Fairwind Program.

What Is Gemini 4 Carbon?

Gemini 4 Carbon is Google’s latest frontier AI model, developed to handle advanced software engineering and cybersecurity work. Its capabilities include identifying vulnerabilities, validating potential security issues, generating patches and helping developers produce code with fewer security weaknesses.

Unlike models focused primarily on answering questions or completing short coding tasks, Carbon is designed to work through longer, multi-step engineering challenges. Google says internal testing shows that the model can continue working for days, including when it encounters dead ends during a task.

The model is also designed to avoid shortcuts, resist prompt injection attacks and collaborate more effectively with users. These capabilities are particularly relevant to cybersecurity, where an AI system may need to investigate a complicated software problem, verify its findings and develop a suitable fix rather than simply suggest a possible solution.

Google has also highlighted improvements in the model’s transparency, saying Carbon makes unsupported claims half as frequently as Gemini 4 Argon. The company further claims that Carbon is twice as likely as Argon to generate code free of vulnerabilities.

Gemini 4 Carbon’s Key Cybersecurity Capabilities

Google says Carbon introduces improvements across several areas of cybersecurity and Software Development. These capabilities are intended to help security teams identify weaknesses, assess risks and develop reliable fixes.

1. Vulnerability Detection

Carbon is designed to identify security weaknesses in software, including bugs that could expose applications or systems to attacks. Google reported that the model scored 88.2% on SEC-bench Pro, an evaluation focused on finding vulnerabilities in real-world software.

Effective vulnerability detection can help developers and security researchers identify problems before they are exploited. However, the practical value of an AI-generated finding depends on whether the issue can be reproduced, validated and addressed correctly.

2. Automated Security Patching

Finding a vulnerability is only one part of the security process. Developers must also understand its cause and produce a fix that resolves the problem without introducing new weaknesses.

Google says Carbon is better at generating secure code and developing high-quality patches. According to the company, Chrome engineers found that Carbon was ten times better at generating high-quality patches than non-Google models in their evaluation.

This result is a Google-reported comparison, and the supplied information does not specify all the models, testing conditions or criteria used to calculate the difference.

3. Exploit Analysis

Carbon also demonstrated capabilities in evaluating how software vulnerabilities might be exploited. Google reported a score of 54.8% on ExploitGym, a benchmark designed to assess a model’s ability to turn vulnerabilities into exploits.

Such capabilities can support authorised security testing by helping defenders understand how a weakness could be abused. At the same time, exploit-generation capabilities require safeguards because the same technical knowledge could be misused against vulnerable systems.

4. Resistance to Prompt Injection

Google says Carbon has been improved to resist prompt injection attacks. These attacks attempt to manipulate an AI system through instructions embedded in user inputs, documents or other content it processes.

Improved resistance is important for cybersecurity applications, where models may examine untrusted code, files or data. Google has described stronger safeguards as part of its approach to deploying Carbon, although the supplied announcement does not provide detailed measurements of its prompt injection resistance.

5. Long-Running Engineering Tasks

Carbon is designed to complete complicated tasks involving multiple stages of investigation, coding and validation. According to Google’s internal tests, it can persist through dead ends and continue working on difficult assignments for days.

This capability could help security teams handle investigations that require repeated testing and adjustments. Google says Carbon also improves on Argon in long-horizon tasks, where an AI system must maintain progress towards a goal over an extended period.

Gemini 4 Carbon Benchmark Results

Google highlighted several benchmark results to demonstrate Carbon’s performance in general AI capabilities, vulnerability detection, exploit analysis and software patching.

These figures were reported by Google in connection with the model’s launch. The supplied information does not provide the complete testing methodology or all comparison scores, so the results should be understood within the context of the specific evaluations.

Performance on the Artificial Analysis Intelligence Index

Carbon secured the top position on the Artificial Analysis Intelligence Index version 5.0, recording a score of 59, according to Google. Claude Opus 5.5 and GPT-6 Astra were reported to be behind it on the index.

The ranking offers an indication of how Carbon performed across the evaluations included in the index. However, an overall benchmark position does not necessarily establish that a model will be the most effective option for every coding, security or business application.

Results on SEC-bench Pro and ExploitGym

Carbon’s reported score of 88.2% on SEC-bench Pro highlights its performance in identifying software vulnerabilities. Its 54.8% result on ExploitGym measures a different capability: determining how vulnerabilities can be turned into exploits.

These evaluations examine different stages of cybersecurity work. Finding a weakness, demonstrating its exploitability and producing a reliable patch are distinct tasks, and performance in one area does not automatically establish equivalent performance in another.

Google also reported a score of 71.0% on CWE-bench for patching. Taken together, the results suggest that the company is positioning Carbon as a model for multiple stages of the vulnerability management process, from discovery to remediation.

Improvement on WebDev Arena

Google said Carbon increased its score by 89 points on WebDev Arena, placing it fourth on the leaderboard. The result adds to the company’s claims that Carbon improves on Argon in software development tasks beyond cybersecurity alone.

Benchmark results should still be interpreted carefully. As Google’s announcement acknowledged, these evaluations offer only a partial picture of model performance, particularly when competing frontier models achieve results that are close to one another.

How Gemini 4 Carbon Compares With Gemini 4 Argon

Gemini 4 Carbon follows Argon, which Google released approximately two weeks earlier. Both models are part of the company’s frontier AI development efforts, but Carbon places a stronger emphasis on cybersecurity and complex engineering tasks.

Google identifies several improvements in Carbon compared with Argon:

• Better vulnerability detection: Carbon is designed to find software security weaknesses more effectively.
• Improved secure coding: Google says Carbon is twice as likely to generate code free of vulnerabilities.
• Fewer unsupported claims: Carbon reportedly cuts unsupported claims by half compared with Argon.
• Stronger long-horizon performance: The model is designed to persist through difficult, multi-step tasks for extended periods.
• Enhanced security safeguards: Google says it has strengthened protections as part of Carbon’s release.

These improvements reflect Google’s stated evaluation results. The company has not provided enough information in the supplied announcement to independently establish how these differences would translate into performance across all real-world applications.

Why Is Google Limiting Access to Gemini 4 Carbon?

Google is taking a gradual approach to releasing Carbon because of its advanced cybersecurity capabilities. At launch, the model is available only to a select group of cybersecurity defence organisations, with broader access planned through the company’s Fairwind Program.

The restricted rollout is intended to give trusted cyber defenders access before wider availability. Google has described the approach as deliberate and responsible, with stronger safeguards incorporated into the model.

Cybersecurity AI presents a particular challenge because capabilities that help defenders find and fix vulnerabilities can also be used to investigate or exploit weaknesses in systems without authorisation. Limiting initial access allows Google to begin deployment with selected partners rather than immediately opening the model to all users.

Google has not provided a specific date for general availability in the supplied announcement. The company said it plans to expand access to additional partners before making Carbon more widely available.

Gemini 4 Carbon Pricing

At launch, Gemini 4 Carbon carries the same introductory pricing as Gemini 4 Argon. Google has set the price at $2 per million input tokens and $10 per million output tokens.

Input tokens generally represent the information sent to a model, while output tokens represent the content it generates. The overall cost of using Carbon will therefore depend on how much information a task requires the model to process and how much output it produces.

The announced rates are introductory prices. The supplied information does not specify whether the pricing will change when access expands or whether additional conditions will apply to different users and partners.

Google’s AI Launch Comes Alongside Its Latest Earnings Report

Gemini 4 Carbon’s debut coincided with Google’s latest earnings report, which showed strong third-quarter growth in its cloud computing business but weaker growth in search advertising revenue.

The company’s stock fell nearly 10% in after-hours trading following the report. The earnings performance and market reaction provide broader business context for the model’s launch, although the supplied information does not establish that Carbon’s announcement caused the share-price movement.

The launch also reflects Google’s continued efforts to develop its frontier AI models across coding, agentic tasks and cybersecurity. By introducing Carbon shortly after Argon, the company is signalling a rapid development schedule for its latest models.

What Gemini 4 Carbon Could Mean for Cybersecurity

If Carbon performs as Google reports in practical environments, it could help security teams automate parts of vulnerability discovery, software testing and patch development. Its ability to work on longer tasks may also prove useful for investigations that require repeated attempts and multiple stages of validation.

However, benchmark performance alone does not establish how reliably the model will operate across different software systems or under real-world attack conditions. Security teams would still need to verify findings, test generated patches and ensure that fixes do not introduce new problems.

Access restrictions will also shape the model’s initial impact. Since Google is starting with selected cybersecurity defenders, its early availability will be narrower than that of a generally accessible AI assistant.

For now, Carbon represents Google’s latest effort to apply frontier AI to defensive security. Its reported results point to improvements in vulnerability detection, exploitation analysis and patching, while its gradual rollout reflects the risks associated with releasing powerful cybersecurity capabilities.

Frequently Asked Questions (FAQs)

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google’s frontier AI model designed to support cybersecurity and software engineering tasks, including vulnerability detection, secure code generation and software patching.

What can Gemini 4 Carbon do?

According to Google, Carbon can identify software vulnerabilities, help generate security patches, analyse potential exploits, resist prompt injection attacks and work through complex, multi-step engineering tasks.

How much does Gemini 4 Carbon cost?

Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens, matching the initial pricing of Gemini 4 Argon.

Is Gemini 4 Carbon available to everyone?

No. At launch, Google is providing access to selected cybersecurity defence organisations and plans to expand availability through its Fairwind Program before a wider release.

How does Gemini 4 Carbon compare with Gemini 4 Argon?

Google says Carbon improves on Argon in vulnerability detection, secure coding, long-running engineering tasks and resistance to prompt injection. The company also reports fewer unsupported claims and a higher likelihood of generating code free of vulnerabilities.

Did Gemini 4 Carbon outperform other AI models?

Google reported that Carbon outperformed GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in almost every test it highlighted. These are company-reported results, and benchmark performance does not necessarily predict performance in every real-world task.