Gemini 4 Carbon puts Google on top of a major index, and its cyber edge is being released only to defenders first

Google has released Gemini 4 Carbon, its latest frontier model. The company claims "significant gains" in coding, agentic tasks and cybersecurity. Google's announcement highlights first place on the Artificial Analysis Intelligence Index (v5.0), a broad composite benchmark, ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra. Carbon succeeds Gemini 4 Argon, which Google shipped earlier this month. That means two frontier releases in a matter of weeks.

The security results are the headline. Google says Carbon ranks first on SEC-bench Pro, which measures whether a model can find real-world bugs, and on ExploitGym, which measures whether it can turn a vulnerability into a working exploit. Google's own chart shows Carbon beating GPT-6 Astra on SEC-bench Pro. Google calls this a "significant leap" in cyberdefense: better at finding vulnerabilities and writing secure code. For now, access is limited to a select group of cybersecurity defense companies. Google says it will later expand to other partners in its Fairwind Program.

Google makes several other claims. Its internal teams have seen Carbon "run for days" to push through dead ends on complex, multi-step engineering work. The company says it takes fewer shortcuts, resists prompt injection better (attacks that hide malicious instructions inside the content a model reads) and works better with users. Google also says Carbon makes half as many "unsupported claims" as Argon. Pricing matches Argon's introductory rate: $2 per million input tokens and $10 per million output tokens.

Our reading: this is a real challenge to one of our standing positions. In our tracking, Google has been solid but behind Anthropic and OpenAI on the hardest tasks. If independent testing confirms a top AA Index score plus first place on SEC-bench Pro and ExploitGym, that gap has closed, at least for now. But we keep our rule: measurement over marketing. Most of these figures come from Google's announcement. A composite index can reward breadth over depth. And a lead in one release cycle can flip within weeks, as Argon-to-Carbon shows. We will judge Carbon on independent, unguided expert tasks, not launch charts.

The release strategy deserves more attention than the leaderboard. ExploitGym measures offensive capability by definition: turning bugs into exploits. A model that leads it is a dual-use tool, useful to defenders and attackers alike. Giving defenders access first is the right instinct. It gives them time to patch before similar capability spreads. On the same day, a startup claimed its AI agents breached a frontier lab through a third-party vendor in under 12 hours. Both stories make the same point. AI cyber capability is now a risk to govern, not a feature to rank.

The long view is still hopeful. Models that find bugs faster than people can write them could eventually make software much safer, just as agentic AI makes attacks cheaper in the short term. Whether defenders or attackers come out ahead depends on who gets access first and under what rules. Google's gated launch is a sensible template. The open question is whether it holds once commercial pressure and the next rival release arrive.