Google Announces Gemini 4 Carbon to Enhance Cyber Defense and Long Term Development
October 31, 2026 · 2 min read
Also published in Türkçe, Deutsch, Français, Norsk, Polski, עברית, Português (Brasil), Svenska, Nederlands
Google announced the Gemini 4 Carbon AI model on October 28, 2026. The model is designed to improve capabilities for cyber defense, specifically in discovering and fixing vulnerabilities, as well as handling long term development tasks that span multiple processes.
Built upon Gemini 4 Argon, which was released on October 13, Gemini 4 Carbon features improved coding and tool usage abilities. Internal Google tests showed the model could spend several days investigating and fixing bugs in low level systems and large codebases, continuing work even when encountering obstacles.
Google stated that the model is better at following project constraints and completing tasks without deviating from the requested scope or failing to perform necessary work. When instructions are ambiguous, the model is designed to research related materials and notify the user of any discrepancies before execution.
In internal testing, the model reduced unfounded claims by approximately half compared to Argon. It is designed to specify limitations in research or verification methods and postpone processes where execution conditions are not met, noting what is missing. On Terminal-Bench 4.0, the model increased the correct answer rate while reducing the successful cheating rate to 0.0 percent and the attempted cheating rate to 0.4 percent.
According to the Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon ranked first with a score of 59, compared to a score of 50 for Argon. Internal benchmarks also indicated that the success rate for generating secure code more than doubled compared to Argon.
Google reported improved resistance to indirect prompt injection. In the Gray Swan IPI benchmark, with 15 attack attempts, the attack success rate dropped from 0.7 percent for Argon to 0.0 percent for Carbon.
The model is being released first to select cyber defense personnel and will be expanded to trusted partners through the Fairwind Program. This program targets government agencies, medical institutions, and telecommunications carriers via a screening process to ensure defensive capabilities are prioritized over potential offensive misuse. Google intends to eventually provide the model to developers, companies, and general users.
Introductory pricing is set at 2 dollars per 1 million input tokens and 10 dollars per 1 million output tokens, with cached input discounted by 95 percent. After the introductory period, prices will increase to 4 dollars for input and 20 dollars for output.