Google has announced Gemini 4 Carbon, a new model based on the Argon architecture and designed to significantly improve capabilities in the cyber domain. Specifically, it focuses on finding vulnerabilities within software and automatically writing the necessary patches to fix them.

Gemini 4 Carbon represents the most advanced artificial intelligence model developed by Google in this specific sector. Its main function is to identify weaknesses present in source code and automatically generate the code necessary to resolve them. In a first phase, access will be limited to a selected group of organizations, but it will be progressively extended to partners that are part of the Fairwind Program.
Gemini 4 Carbon: why it matters
Based on Gemini 4 Argon, the new model inherits and further refines the capabilities of the previous platform. Benchmarks published by Google indicate that Gemini 4 Carbon outperforms Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 in almost all tests conducted, including agentic programming, multimodal understanding, and computer use. Performance differences are particularly evident in features related to cybersecurity.
According to data provided by Google, Carbon demonstrates a success rate more than double that of Argon in generating secure code. The Chrome team found that the new model produces patches over 10 times better than those generated by competing models. Gemini 4 Carbon implements the same security measures present in Argon, with some specific improvements.
Argon is vulnerable to the execution of hidden instructions within web pages or documents – a type of attack known as prompt injection – in 0.7% of tested cases. In contrast, Carbon has demonstrated resistance to all attack attempts during trials conducted by Google. Furthermore, the model is designed to be more “aligned,” meaning it adheres strictly to received inputs and does not seek alternative solutions through vulnerabilities or unexpected behaviors.
What changes and what are the effects
In case of doubts or uncertainties, Carbon does not attempt to guess the correct solution but requires user intervention. Currently, access to Gemini 4 Carbon is reserved for a limited number of organizations operating in the cybersecurity sector. Google will evaluate in the future the possibility of extending the use of the model to companies, developers, and private users.
The announcement was published by Google on October 28, 2026.
Source and further reading on Gemini 4 Carbon: original article.
* Content created with the assistance of artificial intelligence systems.
Hardware Ready Ready to Bench?