Google unveils Gemini 4 Carbon, cybersecurity model ranked first in benchmarks

Google announced Gemini 4 Carbon on October 29, a model intended for complex programming and cybersecurity tasks. Unlike other recent launches by the company, initial access is not open to the public: it is reserved for a group of cyber defenders and partners selected through the Fairwind program, with an extension to developers and businesses planned for later.

According to the Artificial Analysis Intelligence Index, an independent index measuring the performance of large language models, Gemini 4 Carbon reached first place with a score of 59, surpassing the models from OpenAI and Anthropic that had led the ranking for most of the year. The overtaking involves two competing companies simultaneously, a fact that signals how much the large language model sector remains subject to frequent leadership changes measurable through public parameters.

Google has also stated that the model achieved performance more than ten times superior to that of another major competing model in writing high-quality security fixes. This, in this case, is an assessment provided by the company itself and not by the independent index cited previously: a distinction worth maintaining, since the two sources follow different criteria and non-overlapping measurement methodologies.

The choice to reserve initial access to a limited number of cyber defenders, rather than opening it immediately to the public, signals an approach different from the launches of general-purpose models: Google is tying the rollout of the tool to an early-adopter selection program, whose extension to developers and businesses remains, for now, without a precise date indicated in the announcement. The Fairwind program thus remains, at the time of publication, the only declared access channel to the new model.