Google unveils Gemini 4 Carbon model

Google Launches Gemini 4 Carbon Model

Google says Gemini 4 Carbon can spot 88.2 percent of real software bugs, topping the SEC-bench Pro leaderboard. The model also leads ExploitGym with a 54.8 percent exploitation score and hits 71.0 on CWE-bench for automatic patching. Carbon ranks first on the new v5.0 Artificial Analysis Intelligence Index with a score of fifty-nine. It also jumps 89 elo points to fourth place (1769) on WebDev Arena. Built on Gemini 4 Argon released just two weeks earlier, Carbon doubles secure-code generation rates. It can run for days on complex, multi-step engineering tasks. Internal tests show misalignment alerts fell by eighty percent and cheating attempts dropped to zero on Terminal-Bench 4.0. Google will roll it out this week to a select group of cyber defenders through the Fairwind Program. Access is planned later. Pricing mirrors Argon at two dollars per million input tokens and ten dollars per million output tokens. Cached inputs are priced at ninety-five percent off. The Chrome team reported Carbon performed more than ten times better than another frontier model when generating quality patches. At an 88.2 percent success rate on SEC-bench Pro, Carbon sets a new benchmark for AI-driven cybersecurity.