Gemini 4 Carbon Introduced as Advanced Cybersecurity Model

On October 28, 2026, Gemini 4 Carbon was announced as the newest frontier model, boasting advanced cybersecurity features and long-horizon capabilities.

Carbon has achieved the top ranking on the updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59. This model shows significant improvements in coding and agentic tasks, particularly in cybersecurity defense, leading in SEC-bench Pro (88.2%) for vulnerability detection, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching. Internal benchmarks indicate that Carbon enhances secure code generation and reduces misalignment and fabrication rates in live coding environments.

Building on the recently launched Gemini 4 Argon, Carbon represents a rapid advancement in capabilities. The development team is focused on enhancing model performance while strengthening safeguards.

The rollout of Carbon will be gradual, starting with a subset of cyber defenders as part of the U.S. government’s voluntary pre-release model testing process. Access will be expanded to additional trusted partners through the Fairwind Program, with plans to gather feedback and refine guardrails before making the model available to developers, businesses, and consumers.

Gemini 4 Carbon will be offered at the same introductory price as Argon, set at $2 per million input tokens and $10 per million output tokens, with cached input tokens available at a 95% discount.

Enhanced Performance for Complex Tasks

Designed for challenging engineering problems, Carbon supports sustained agentic workflows. It has demonstrated the ability to run for extended periods, addressing complex, multi-step tasks such as debugging and patching across large codebases.

Improvements in model performance were made based on user feedback, focusing on coding, agentic, scientific tasks, full-stack web development, and multimodal understanding. Carbon’s output quality has improved, resulting in a jump of 89 elo points to a score of 1768 on the WebDev Arena leaderboard.

Cybersecurity Advancements

Gemini 4 Carbon delivers a notable enhancement in cybersecurity capabilities, achieving state-of-the-art performance across the cybersecurity defense lifecycle, including vulnerability detection, validation, and remediation.

Internal benchmarks show that Carbon has more than doubled the success rate for generating secure code compared to Argon. The Chrome team reported that Carbon outperformed another leading model by over 10 times in generating high-quality patches. Additionally, Wiz, a cloud security platform, noted Carbon’s superior performance in web penetration testing.

Key external benchmarks highlight Carbon’s capabilities:

• Finding real vulnerabilities: Carbon achieved 88.2% on SEC-bench Pro, effectively uncovering complex flaws in real-world software.
• Validating risk through exploitation: Carbon set a new record on the ExploitGym benchmark with a score of 54.8%.
• Patching at the root cause: Carbon reached 71.0% on CWE-bench, translating validated findings into effective code fixes.

Focus on Safety and Security

Gemini 4 Carbon builds on the safety and security measures of Argon, incorporating model training, real-time misuse defenses, and prompt-injection resilience. Significant improvements in model alignment have been made, with a focus on appropriate responses to complex tasks.

Carbon has reduced successful shortcut attempts during training to 0.0%, while also improving overall accuracy. In live coding scenarios, high-severity misalignment alerts decreased by nearly 80%.

Transparency has also improved, with unsupported claims cut by approximately 50% compared to Argon. Carbon provides clear methodological caveats and avoids fabricating information.

Furthermore, Carbon enhances collaboration with users by retrieving relevant sources and addressing discrepancies before proceeding with actions.

With a commitment to safety and reliability, Gemini 4 Carbon aims to deliver significant advancements in software engineering and cybersecurity defense. The model is expected to be available to developers, businesses, and consumers in the near future.

Source: Google DeepMind