Google Splits AI Roles With 'Carbon' for Long-Haul Tasks

According to AI TIMES, Google unveiled a new frontier model, "Gemini 4 Carbon," on November 10 that is specialized for complex long-horizon tasks and repeated verification, outlining a strategy that splits general-purpose work models from high-level reasoning models. The model ranked first in three categories of agentic AI evaluations, and its real-world task costs were also said to be less than half those of comparable models. The key point was not its performance alone, but that it was designed for work that runs through multiple steps over a long period, fixing errors and checking results again, such as software development, security, and scientific research. Google also introduced a "swarm" in which multiple AIs divide roles and collaborate, emphasizing a structure that moves away from the approach in which a single model handles the entire process. In security, its use cases were expanded to vulnerability detection and patch support, and in science, to data interpretation and hypothesis exploration, but the article also noted that these figures and examples came from a specific evaluation environment. In the end, the announcement was read as a signal that the benchmark for frontier AI competition is shifting from simple answer quality to long-task completion rates, verification ability, and cost efficiency.