Google launches Gemini 4 Carbon for cybersecurity and coding

Introducing Gemini 4 Carbon

Google has officially rolled out the Gemini 4 Carbon series, a new generative artificial intelligence model designed for high-level performance. This release follows the recent launch of the Gemini 4 Argon model earlier this month. While Argon introduced advanced generative capabilities, Carbon is positioned as a more specialized tool for complex tasks.

The Carbon model is built directly on the Gemini 4 Argon architecture. Google describes it as its strongest model to date, specifically highlighting its state-of-the-art capabilities in cybersecurity and long-horizon tasks. This means the model is built to maintain focus and progress over extended periods of time without losing track of the original goal.

According to the Deccan Herald, the model has already secured a top spot on the Artificial Analysis Intelligence Index v5.0. It received a score of 59 on this recently updated index, which measures the intelligence and capabilities of various AI models. This ranking places Carbon at the forefront of current large-scale language models.

Cybersecurity and coding benchmarks

A primary focus of the Gemini 4 Carbon model is its ability to handle security-related tasks. It has shown significant gains in areas critical for defense, such as finding vulnerabilities and patching them. In the SEC-bench Pro test, the model achieved a score of 88.2% for finding vulnerabilities, while it scored 71.0% on the CWE-bench for patching.

The model also performed well in exploitation tasks, scoring 54.8% on ExploitGym. These benchmarks are used to measure how well an AI can understand the mechanics of cyberattacks and the methods required to stop them. Google aims to use these capabilities to help identify hard-to-detect security flaws within various applications.

In practical testing, the Chrome team reported that Carbon outperformed another lab's large frontier model by more than 10X when generating high-quality secure patches. Additionally, the security platform Wiz observed that Carbon led on benchmarks for web penetration testing. This is a core part of their Scan for Good initiative, which looks for security weaknesses across the internet.

Agentic AI and long-horizon tasks

One of the most distinct features of Gemini 4 Carbon is its ability to function as an agentic AI. This means the model is programmed to work independently for days to complete a single, complex job. It is designed to persist through dead ends and navigate multi-step engineering problems without constant human intervention.

The model is specifically tuned to handle tasks like root-causing and patching subtle bugs across very large codebases and low-level systems. It can understand complex queries and maintain its trajectory on difficult jobs. This capability is intended to assist with engineering tasks that require deep reasoning and many sequential steps to reach a conclusion.

Internal testing by Google's own teams showed that the model delivers a strong uplift in generating secure code. These tests also revealed that the model showed reduced rates of fabrication and misalignment in live coding environments. Fabrication refers to when an AI generates false information or hallucinated facts during a task.

Pricing and availability

Google is making the Gemini 4 Carbon model available to developers, businesses, and consumers at an introductory price point. The pricing structure is designed to match the previous Gemini 4 Argon model. This ensures that users can access the high-end capabilities of Carbon without a significant jump in costs compared to the previous generation.

The cost for using the model is set at $2 per million input tokens and $10 per million output tokens. Tokens are the basic units of text that the AI processes. For users who need to reference the same information multiple times, Google is offering a discount on cached input tokens.

Specifically, cached input tokens are priced at 95% off the standard input token price. This pricing model allows for more efficient and cost-effective use of the model when dealing with large amounts of data. The model is now ready for integration into various workflows that require high-level coding and security analysis.