Google launches Gemini 4 Carbon for cybersecurity and long-term coding

A new frontier for security

Google has officially introduced Gemini 4 Carbon, a new frontier AI model specifically designed to excel in cybersecurity and autonomous software engineering. The model arrives just two weeks after the release of Gemini 4 Argon and currently holds the top spot on the Artificial Analysis Intelligence Index v5.0. It is built to address high-level technical needs that require more than just simple text generation.

One of the primary focuses for Carbon is its ability to identify and repair software vulnerabilities. According to the source, the model achieved an 88.2% score on SEC-bench Pro for finding flaws and a 71.0% score on CWE-bench for creating security patches. These benchmarks measure how well an AI can spot weaknesses in code and suggest ways to fix them before they can be exploited.

In internal testing, Google reported that Carbon more than doubled the success rate of the Argon model when generating secure code. While these external evaluations were conducted without safety guardrails, they provide a baseline for the model's raw capabilities. The company is now working to establish how these results translate into real-world production environments.

Handling long-term engineering

Unlike many current AI tools that provide quick answers to short prompts, Carbon is designed to manage engineering assignments that last for several days. This allows the model to work through complex tasks like diagnosing difficult software bugs across very large codebases. It is intended to function as an autonomous agent that can persist through a project until completion.

Google's internal teams have already begun using the model to resolve difficult software issues with less human intervention than previous tools. During these tests, the model demonstrated the ability to continue working even when it encountered unsuccessful attempts to finish a specific task. This persistence is a key differentiator for the model's role in enterprise technology operations.

The model's performance in coding was also reflected in the WebDev Arena, where it saw an 89-point Elo improvement. This metric helps rank the model's ability to complete web development tasks compared to other systems. By staying on a task for longer periods, Carbon aims to reduce the manual oversight typically required for large-scale software maintenance.

Reliability and safety guardrails

To make the model viable for professional use, Google focused on reducing unwanted behaviors and improving reliability. Internal evaluations showed that Carbon had an almost 80% reduction in high-severity misalignment alerts when compared to the Argon model. Misalignment refers to instances where the AI's output does not align with the user's intended goals or safety standards.

The company also reported that Carbon produces fewer unsupported claims and shows improved resistance to prompt injection. Prompt injection is a technique where a user tries to trick an AI into ignoring its instructions or performing prohibited actions. Carbon successfully resisted every attack included in the evaluated Gray Swan benchmark during these tests.

Despite these improvements, Google is being cautious about how the model is deployed. The company is currently limiting initial access to selected cybersecurity defenders and trusted partners. This restricted rollout allows for continued safety testing and feedback gathering before the model is made available to a broader audience.

Deployment and pricing details

Google is currently participating in voluntary U.S. government pre-release testing to further refine the model's safeguards. The company plans to expand access to the general public once they have gathered enough feedback from their initial partners. However, Google has not yet announced a specific date for when the model will be generally available.

The pricing for Gemini 4 Carbon will match the introductory rates set for the previous Gemini 4 Argon model. The starting costs are $2 per million input tokens and $10 per million output tokens. For users who need to process large amounts of data, cached inputs will receive a 95% discount.

While these figures provide a baseline for usage, Google has not yet provided detailed production cost estimates for large enterprise deployments. This information will likely be necessary for organizations looking to integrate Carbon into their existing infrastructure. For now, the focus remains on refining the model's ability to handle autonomous tasks safely.