Google announced Gemini 4 Carbon on October 28, introducing its new frontier model through an initial rollout limited to a group of cyber defenders. The company describes it as its strongest model yet for cybersecurity and long-horizon tasks, but it is not generally available to developers, businesses, or consumers. The launch pairs high benchmark results with a cautious distribution strategy: phased access, pre-release testing with the U.S. government, and feedback collection before broader availability. [1]

According to Google, Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The company also reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for vulnerability patching. These results matter because they cover several stages of defensive work — finding, validating, and remediating flaws — although published benchmark numbers do not by themselves establish performance on production networks. Some of the evidence comes from internal evaluations and claims made by Google teams. [1]

Carbon follows Gemini 4 Argon, which Google introduced only two weeks earlier. Google says the newer model improves on coding, agentic tasks, and secure-code generation. It also says Carbon showed lower rates of misalignment and fabrication in live coding environments used by its own teams. The announcement does not provide every element needed for outside researchers to reproduce each comparison, so the reported metrics should be distinguished from comprehensive independent validation.

The initial access plan reflects that risk. Google said it is participating in the U.S. government's voluntary process for testing models before release and that Carbon will begin reaching a subset of cyber defenders this week through the Fairwind Program. Access will expand gradually to additional trusted partners while the company iterates on safeguards. This is therefore neither a general release nor evidence that the model is already operating broadly across critical organizations. [1]

Ars Technica characterized Carbon as more of an improvement over Argon than an entirely separate model family. Its coverage also highlights the central tension in the announcement: Google is promoting significant advances, particularly in security, while most users still cannot test them. The publication cites additional company claims, including a greater likelihood of producing vulnerability-free code and improvements in patch generation, but presents them as results reported by Google. [2]

Restricted access is consequential because the same capabilities that help defenders locate and repair flaws can increase a model's offensive potential. Google says it is strengthening safeguards while accelerating development. At this stage, the most concrete signal is not only Carbon's position on a leaderboard but also its distribution model: selected defenders first, then trusted partners, followed later by products and API access for a broader audience.

Gemini 4 Carbon raises the level of competition around models that can complete extended technical work and act on code rather than merely answer questions. It also raises the evidentiary standard these announcements should meet. Independent results, details about practical access, and the model's behavior on defensive tasks outside company laboratories will determine whether the reported advantages translate into operational improvements. For now, Carbon is a confirmed, high-impact launch, but one with controlled availability and several performance claims that still await broad external validation. [1][2]

Sources

[1] Read More — Gemini 4 Carbon: our next frontier model — Google
[2] Read More — Google announces Gemini 4 Carbon, its latest frontier AI model that you can't use — Ars Technica