Google Launches Gemini 4 Carbon With Cybersecurity Focus

Google has launched Gemini 4 Carbon cybersecurity capabilities as a central part of its latest frontier AI model, stepping up competition with OpenAI and Anthropic in advanced coding and security tools. The model is designed to identify software vulnerabilities, validate potential exploits and help developers patch weaknesses, while handling complex engineering tasks that can run for extended periods.

Announced on October 28, Gemini 4 Carbon follows Google’s Gemini 4 Argon release by about two weeks. Google is initially making Carbon available to a limited group of cybersecurity defenders as it continues testing the model’s safeguards before a broader rollout.

Gemini 4 Carbon Cybersecurity Capabilities Target Vulnerabilities

Google says Carbon delivers strong performance across several stages of the cybersecurity process, including vulnerability discovery, exploit validation and software remediation. The goal is to help security teams move from identifying a weakness to understanding whether it can be exploited and developing a fix.

On Google’s internal benchmarks, Carbon generated secure code with more than twice the success rate of Gemini 4 Argon. The company also reported a score of 88.2% on SEC-bench Pro for finding vulnerabilities and 71% on CWE-bench for patching them. These results reflect Google’s published evaluations and should not be treated as proof that the model will deliver the same performance across every real-world system.

The model is also designed for long-running, multi-step engineering work. Google says its internal teams have used Carbon on complex tasks such as tracing subtle bugs across large codebases and developing fixes. These capabilities could be useful for security researchers and software teams managing extensive applications, although their practical value will depend on testing in real deployment environments.

Google Limits Initial Access While Testing Safeguards

Google is taking a staged approach to releasing Gemini 4 Carbon. Initial access is being offered to a subset of cybersecurity defenders through its Fairwind Program, with additional trusted partners expected to receive access gradually.

The company said it is continuing to gather feedback and refine safeguards before making the model available more widely to developers, businesses and consumers. It is also participating in the US government’s voluntary process for testing frontier models before release.

The cautious rollout reflects a challenge for advanced AI security tools: capabilities that help defenders find and repair weaknesses may also be useful to people attempting to exploit them. Limiting early access gives Google an opportunity to assess how the model performs and how its protections hold up before broader deployment.

Google says Carbon builds on the safety and security measures introduced with Argon, including model training, real-time misuse defences, prompt-injection resistance and monitoring. The company has reported improvements in some internal safety evaluations, but independent testing will be important in assessing how reliably those protections work outside controlled environments.

Gemini 4 Carbon Pricing and Availability

During its introductory pricing period, Gemini 4 Carbon is priced at $2 per million input tokens and $10 per million output tokens, matching Gemini 4 Argon’s introductory rates. Google says those prices will later increase to $4 per million input tokens and $20 per million output tokens. Cached input tokens receive a separate discount.

The initial release is aimed at cybersecurity professionals rather than the general public. Google has not announced a firm date for wider availability, saying it intends to expand access as testing and safeguards progress.

AI Competition Moves Further Into Cybersecurity

The launch shows how major AI developers are competing not only on general reasoning and coding, but also on tools that can help organisations secure software. For companies facing growing codebases and increasingly complex infrastructure, models that can find vulnerabilities and assist with repairs may become part of security workflows.

However, benchmark results alone cannot establish how much a model reduces risk in production. Reliability, false positives, the quality of generated fixes and the ability to resist misuse will all matter as Gemini 4 Carbon moves beyond its initial group of testers.