Google launches Gemini 4 Carbon to help identify software vulnerabilities

A new frontier for security

Google has officially introduced Gemini 4 Carbon, a new artificial intelligence model built to tackle complex software security challenges. The company is positioning this model as a tool for finding vulnerabilities, testing potential exploits, and developing actual fixes for software. Because of the sensitive nature of these capabilities, Google is initially limiting access to cybersecurity defenders while it works to strengthen the model's safeguards.

CEO Sundar Pichai announced the launch on social media, noting that the release follows the recent debut of Gemini 4 Argon. He highlighted that the rapid succession of these releases demonstrates the pace at which the company is pushing the boundaries of AI capabilities. Carbon is intended to be a frontier model, meaning it represents the current state-of-the-art in what these systems can achieve.

The model's performance has already been validated by external benchmarks. According to Google, Gemini 4 Carbon ranked first on the v5.0 of the Artificial Analysis Intelligence Index. It earned a score of 59 on this index, which measures the intelligence and capabilities of various frontier models. This ranking places it at the top of its class for the specific tasks it was designed to perform.

How the model handles code

Google DeepMind designed Carbon to handle engineering tasks that require looking at code over extended periods. Unlike simpler models that might only look at a single block of text, Carbon can trace bugs back to their root causes. This is particularly useful for large codebases and low-level systems where a single error might be buried deep within thousands of lines of instructions.

The technical testing for the model shows significant progress in specific security categories. It scored 88.2 percent on SEC-bench Pro, which is a benchmark used for identifying software vulnerabilities. Additionally, it scored 54.8 percent on ExploitGym, a tool used to test whether discovered vulnerabilities can actually be exploited by malicious actors.

When it comes to fixing the problems it finds, Carbon scored 71 percent on CWE-bench, which evaluates the development of fixes. The company's Chrome team reported that the model was significantly more effective at creating high-quality patches than a competing frontier model from a different AI lab. These results suggest that the model is becoming increasingly proficient at the actual repair work required by developers.

Safety and accuracy gains

A major goal for Google in developing Carbon was improving the reliability of the AI's outputs. Internal tests showed that Carbon achieved a roughly 50 percent reduction in unsupported claims when compared to the previous Argon model. This means the model is less likely to make up facts or provide incorrect information when asked to perform complex coding tasks.

Security and alignment are also central to the model's development. Google reported a nearly 80 percent decline in high-severity misalignment alerts within coding environments. Misalignment refers to instances where an AI's behavior deviates from the intended safety guidelines or human instructions. Reducing these alerts is a key step in making the model safer for professional use.

The model also demonstrated resilience against common AI manipulation techniques. It recorded zero successful cheating attempts on Terminal-Bench 4.0. Furthermore, it showed no successful attacks on the Gray Swan indirect prompt-injection benchmark during evaluation. Prompt injection is a method where users try to trick an AI into ignoring its instructions to perform prohibited actions.

Rollout and pricing details

Google is taking a measured approach to making Gemini 4 Carbon available to the public. While the model is technically ready, the company is rolling it out gradually and responsibly. Some trusted cyber defenders are receiving access starting this week as a first step in the deployment process.

The company plans to expand access to more partners through its Fairwind Program before a wider release. This phased approach allows Google to monitor how the model performs in real-world security environments. It also provides a window for the company to ensure that the model's capabilities are being used for defensive purposes rather than harmful ones.

For those who eventually gain access, the pricing will remain consistent with the previous Argon model. The cost is set at $2 per million input tokens and $10 per million output tokens. To help with efficiency, cached input tokens will receive a 95 percent discount, which can lower the cost of repeated queries on the same data.