Google Introduces Gemini 4 Carbon for Programming and Cybersecurity Defense
Google announced the Gemini 4 Carbon model, focused on programming and cybersecurity defense. It is currently available only to selected defenders and partners through the Fairwind Program.
On October 28, Google announced Gemini 4 Carbon, a new frontier model focused on programming and cybersecurity defense tasks. The model is not yet generally available: the company is deploying it only to selected cyber defenders and trusted partners through the Fairwind Program.
Google did not specify when the model might become available to a broader group of developers, companies, or consumers. The immediate practical impact of the announcement is therefore limited to organizations involved in the initial testing and deployment.
Gemini 4 Carbon in a Controlled Rollout
According to Google, the phased availability is part of its approach to a model whose capabilities are intended to help find, verify, and fix software bugs. The company says it wants to continue testing safeguards and gather feedback from the Fairwind Program before broader deployment.
This restriction is also significant given the nature of the stated tasks. Capabilities that can be used to analyze vulnerabilities and create more secure code may support defensive workflows, but Google is currently making them available only to a controlled group of users.
The announcement itself does not demonstrate the actual exploitation of a security flaw or an attack carried out using this model. Google describes evaluation results and planned use in cybersecurity defense.
Pricing and Published Evaluations
Google set an introductory price of $2 per million input tokens and $10 per million output tokens. After the initial period, prices are expected to rise to $4 per million input tokens and $20 per million output tokens.
The independent service Artificial Analysis lists a score of 59 for the Gemini 4 Carbon (High) variant and first place in the Intelligence Index v5 ranking. This is an evaluation based on that specific methodology, not a universal assessment of the model’s capabilities across all types of tasks.
Google also published the results of its own security benchmarks: 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. These figures come from company materials and were not independently reproduced in the available documentation.
The company also claims that, when generating secure code, the model achieved more than a twofold improvement over Gemini 4 Argon in internal benchmarks. Here too, this is a company measurement, and its results should be distinguished from independent tests.
What Comes Next
The next important step will be Google’s decision on broader availability through APIs, Gemini services, or enterprise products. Independent reproductions of the security benchmarks, any additional technical documentation, and the experiences of partners participating in the Fairwind Program will also be monitored.
The announcement places Google in the ongoing competition among models for agentic programming and security workflows. However, while the model remains access-restricted, its stated capabilities cannot be assessed through ordinary public use.
Sources
- Google – Confirms the announcement, limited rollout through the Fairwind Program, stated benchmarks, and pricing.
- Artificial Analysis – Confirms the score of 59 and Gemini 4 Carbon (High)’s first place in the Intelligence Index v5.
- Reuters – Independently summarizes the model’s launch, cybersecurity focus, and limited initial availability.
- Ars Technica – Confirms that general access to the model was not available at the time of the announcement and that the rollout is phased.
Verified and updated: October 29, 2026 06:19
