Gemini 4 Carbon was designed to handle more complex problems than those addressed by Gemini 4 Argon, the model recommended by Google for everyday tasks, software engineering, and professional activities. To do so, the new AI uses a larger number of tokens, with longer reasoning cycles and response verification.
The model also ranked first in three independent benchmarks, each focusing on a different skill: CursorBench 4.0 tests performance in programming and software development; WANDR, from Perplexity, evaluates the ability to conduct in-depth research; and AutomationBench, from Zapier, measures task execution and workflow automation. According to Google, the cost per task came in more than two times lower than that of the closest external competitor.
Google Antigravity also gained a preview of the /swarm command, which coordinates multiple artificial intelligence agents in parallel from a single instruction.The feature can be used to develop code systems, create 3D visualizations, and perform formal security checks.
