Google launches Gemini 4 Carbon for cybersecurity tasks

A specialized tool for defenders

Google has officially announced Gemini 4 Carbon, a model designed specifically for long-horizon tasks. Unlike models that focus on quick interactions, Carbon is built to handle complex engineering work that requires persistence over time. As gHacks reported, the company is launching this model first to a specific group of cyber defenders.

The model arrived just two weeks after the release of Gemini 4 Argon. While Argon serves as a general-purpose model, Carbon is tuned for deeper technical analysis. It is intended to help security teams navigate large codebases to find and fix subtle bugs that might otherwise be missed by human reviewers or simpler AI tools.

The rollout is currently restricted to a subset of cyber defenders through the Fairwind Program. Google is using this period to gather feedback and iterate on the model's safety guardrails. Developers, businesses, and general consumers do not have access to the model yet, but Google plans to open it up as soon as possible.

Performance in security testing

Google released several benchmarks to demonstrate how Carbon handles security-specific tasks. On the SEC-bench Pro, which measures the autonomous discovery of real-world vulnerabilities, the model achieved a score of 88.2%. It also scored 71.0% on CWE-bench, a metric that tracks the ability to turn validated flaws into actual working code patches.

In internal testing, the Chrome security team observed significant improvements in patch generation. They noted that Carbon was able to create high-quality security fixes at a speed exceeding ten times that of a large frontier model from a different lab. Furthermore, Google claims Carbon has more than twice the success rate of Argon when it comes to generating secure code in internal benchmarks.

The model also showed strong performance on general coding and web development platforms. It reached a score of 1768 on the WebDev Arena, which currently places it fourth in that ranking. Additionally, it earned 89 Elo points on that same platform to reach its current standing. These scores suggest a high level of competency in practical programming environments.

Comprehensive benchmark rankings

Beyond specific security metrics, Carbon has shown strong performance across broader AI evaluation indices. Google reports that the model currently holds the top spot on the Artificial Analysis Intelligence Index v5.0 with a score of 59. This ranking highlights its capability in handling sophisticated reasoning tasks beyond simple text generation.

The model is designed to handle engineering work that spans multiple days. During internal testing, the model demonstrated the ability to push through various roadblocks to diagnose and root-cause issues. This persistence allows it to handle complex problems that require a deep understanding of how different parts of a large codebase interact with one another.

By focusing on these long-horizon tasks, Google aims to provide a tool that can autonomously navigate systems to find and fix deep-seated bugs. This is a significant step up from models that only provide immediate responses to single prompts. It allows for a more thorough investigation of technical debt and security flaws within large-scale software projects.

Safety and government testing

The model is built on the same safety stack as its predecessor, Gemini 4 Argon. Google reported that during testing, the rate of successful attacks on the Gray Swan indirect prompt injection benchmark reached 0.0% for Carbon. This is a notable improvement over the 0.7% success rate recorded for Argon during the same testing period.

In live internal coding environments, the model also showed a decrease in high-severity misalignment alerts. Google noted that these alerts dropped by almost 80% when using Carbon. These safety metrics are intended to ensure that the model remains reliable and secure while performing autonomous tasks in sensitive environments.

Additionally, Google is participating in a voluntary pre-release model testing process organized by the US government. This participation suggests a commitment to ensuring the model behaves predictably in high-stakes scenarios. The company continues to refine its guardrails while gathering feedback from its initial group of trusted cyber defense partners.

Pricing and availability details

Google has provided specific pricing details to help developers plan their budgets for the upcoming launch. Introductory API pricing matches Gemini 4 Argon, starting at $2 per million input tokens and $10 per million output tokens. To help manage costs, cached input tokens are currently offered at a rate 95% lower than the standard input price.

These introductory rates are temporary. Once the initial period concludes, the price for input tokens will increase to $4 per million, while the price for output tokens will rise to $20 per million. This structure allows early adopters to experiment with the model's capabilities at a lower cost before the standard pricing takes effect.

While the model is currently limited to the Fairwind Program, Google intends to expand access to the wider public soon. The company has not yet announced a specific date for when developers, businesses, and consumers can begin using the API. For now, the focus remains on refining the model's performance and safety based on feedback from the initial security defender wave.