Two weeks after launching Gemini 4 Argon, Google has announced Gemini 4 Carbon, which it calls its strongest model yet, built for cybersecurity and long-horizon engineering work. The announcement, made by Google DeepMind's Koray Kavukcuoglu on 28 October, claims the top position on the Artificial Analysis Intelligence Index at 59 points.
The benchmark numbers explain the confidence. Carbon scores 88.2 percent on SEC-bench Pro, which tests whether a model can autonomously find real vulnerabilities buried in production software. It reaches 71 percent on CWE-bench, which measures the ability to produce working fixes for known weakness categories, and 54.8 percent on ExploitGym, which tests exploitation. It also climbs to fourth place on WebDev Arena for full-stack web development.
What separates Carbon from a typical model release is its intended work. This is not a chatbot upgrade. Google describes teams running it for days at a stretch, persisting through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems. It is designed as an engineering agent, not a conversationalist.
The internal results Google cites are striking. Carbon reportedly more than doubles Argon's success rate at generating secure code. The Chrome team says it produced high-quality patches at more than ten times the rate of another laboratory's frontier model. Wiz, Google's cloud security platform, observes it leading benchmarks for web penetration testing, a capability Wiz uses in its own scanning program.
Yet almost nobody can use it. Starting this week, access goes only to a subset of cyber defenders, with additional trusted partners phased in through Google's Fairwind Program. Google is participating in the United States government's voluntary pre-release testing process. Developers, businesses and consumers must wait.
The restraint is the story. A model that finds and validates vulnerabilities at this level is valuable to defenders and to attackers in equal measure. Google is publicly testing that tradeoff before opening the gates, and the access decisions may reveal more about the model's true capability than the scoreboards do.
The pricing tells the other half of the story. Carbon will launch at the same introductory price as Argon: two dollars per million input tokens and ten per million output tokens, with cached input at a 95 percent discount. This is production economics, not research pricing. Google expects this model to work for a living once released.
For leaders, two lessons follow. First, capability leadership now includes deployment judgment: the most meaningful signal about a frontier model is increasingly who is allowed to touch it and when. Second, security teams should prepare for a world where AI finds flaws faster than organisations patch them, because the defensive version of this capability will soon be standard equipment.
The frontier is no longer just a leaderboard. It is a gate. The organisations that build their security workflows and governance now, before access widens, will be the ones ready when the gate opens.
Think mathematically. Engineer intelligently. Lead strategically. Build responsibly. Scale globally. Create real business value through artificial intelligence.
