Google launches Gemini 4 Carbon for elite cybersecurity teams

A specialized tool for defense

Google officially introduced Gemini 4 Carbon on October 28, 2026. The company describes it as its most capable development to date for cybersecurity and complex long-term tasks. It follows the release of Gemini 4 Argon, which became generally available just two weeks prior.

Unlike general models, Carbon is designed for high-stakes security work. It is intended for activities such as authorized attack simulations and reverse engineering. These processes involve breaking down software to understand how it works or how it might be exploited by bad actors.

The model also excels at analyzing malware, which is software designed to damage or gain unauthorized access to computer systems. Google has placed strict boundaries on its use, explicitly forbidding the creation of new malware or the sale of the model itself. This ensures the tool remains in the hands of defenders rather than attackers.

Benchmarks and performance

In security tests conducted without safety filters, Carbon achieved an 88.2% success rate on the SEC-bench Pro vulnerability benchmark. This specific test measures the ability to identify security flaws across a split of 183 tasks. In comparison, the competitor model GPT-6 Astra reached a score of 85.4%.

The model also performed well in the ExploitGym test, which evaluates the exploitability of security flaws without internet access. Carbon reached a success rate of 54.8% within a ten-hour limit. This outperformed GPT-6 Astra, which scored 42.4% but was not subject to the same time constraints.

Internal testing by Google showed that Carbon more than doubled the success rate for generating secure code compared to Gemini 4 Argon. Additionally, the Chrome team reported that high-quality patches were produced more than ten times more frequently using Carbon than with another leading model. However, the model did see a rise in hallucinations, reaching 11% in the AA-Omniscience test.

Restricted access and partners

Access to Gemini 4 Carbon is currently highly regulated and not available to the general public. Google is initially providing the model to selected security professionals before rolling it out through the Fairwind program. This network includes more than 650 partners globally, prioritizing government agencies and operators of critical infrastructure.

Users of the model must adhere to strict security requirements to maintain access. These include the use of phishing-resistant multi-factor authentication and a requirement that the model only be used by internal defense teams. Continuous monitoring is also required to ensure the tool is used for authorized purposes only.

While the model is currently restricted, Google has not yet announced a date for a wider release to developers or consumers. The current rollout focuses on essential technology platforms and high-priority security partners. This phased approach allows Google to manage the risks associated with such a powerful cybersecurity tool.

Technical specs and pricing

Gemini 4 Carbon is a multimodal model, meaning it can process and understand different types of data simultaneously. It supports text, images, video, and audio inputs. It also features a context window of one million tokens, which allows it to process very large amounts of information in a single session.

The pricing structure for the model's API is set to begin with a discounted rate. At market launch, the price is expected to be $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached inputs. These prices are intended to be the introductory rates for early adopters.

Eventually, the regular price for the API is expected to rise to $4 for input and $20 for output. As of October 30, 2026, the model was not yet active in the Gemini API or the Fairwind catalog. This delay suggests that Google is still finalizing the deployment and safety protocols for the new model.

The competitive landscape

Carbon is part of a growing competition between Google, OpenAI, and Anthropic to dominate the AI space. On the Artificial Analysis Intelligence Index v5.0, Carbon took the first place with 59 points. It narrowly beat out Claude Opus 5.5, which scored 58 points, and GPT-6 Astra, which scored 57 points.

Despite its lead in several categories, Carbon did not win every metric. For example, it fell behind Claude Opus 5.5 in the Coding Agent Index, scoring 59 points compared to Claude's 62. In the OSWorld-2.1 offline test, Claude Opus 5.5 also outperformed Carbon with a score of 87.2% against Carbon's 81.5%.

Google is also participating in voluntary US review processes based on agreements made in May 2026. These steps are part of the broader effort to address compliance risks and legal frameworks like the EU AI Act. The release of Carbon highlights the ongoing push to create specialized models that can navigate the complex demands of enterprise security.