Google's Gemini 4 Carbon tops security benchmarks, but only vetted cyber defenders in its Fairwind Program can use it for now

The reason for the wait is the same thing Google is bragging about. Koray Kavukcuoglu, Google DeepMind’s chief AI architect, lists Carbon’s security scores up front: “SEC-bench Pro (88.2%) for finding vulnerabilities, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching.” A model that finds and exploits bugs that well is useful to whoever holds it.

So Google is choosing who holds it first. The company says it’s “actively engaged in the U.S. government’s voluntary process for pre-release model testing,” and that it will “gradually phase access to additional trusted partners through our Fairwind Program” before a wider release. The Fairwind Program page says Google works “with over 650 partners globally,” prioritizes “governments and national cyber authorities,” critical infrastructure and core tech platforms, and runs “background checks on organizations that apply.” Partners agree to “phishing-resistant MFA,” may only give access to internal security, incident response or pentest teams, and can’t “share, redistribute, or sell access.”

On general smarts, Google says “Carbon ranks #1 on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59.” It climbed “+89 elo points to land at #4 (1768)” on the WebDev Arena coding leaderboard. And “internally, our teams have seen Carbon run for days to persist through dead ends.”

	Gemini 4 Carbon	Gemini 4 Argon
Who can use it	Fairwind cyber defenders first	launched two weeks earlier
Introductory price, per 1M tokens	$2 input, $10 output	$2 input, $10 output
Price after the intro period	$4 input, $20 output	not stated in Carbon’s post
Gray Swan indirect prompt injection, attack success	0.0%	0.7%
Secure code generation	“more than 2x uplift in success rate” over Argon	baseline

All figures come from Google’s Carbon announcement.

Google also leaned on safety numbers. It says Carbon “reduced successful cheating on Terminal-Bench 4.0 to 0.0%,” that “high-severity misalignment alerts dropped by almost 80%,” and that it “cuts unsupported claims by approximately 2x over Argon.” The details are in the Gemini 4 safety report. Google also cites Wiz, which it calls “our leading Cloud security platform,” as seeing Carbon lead on web penetration testing benchmarks. Wiz is Google’s own, so that’s still a house number.

What Google didn’t say matters too. The post gives no context window, no speed figures and no date for general access.

Prediction, high confidence: other frontier labs keep copying this staged release for models with strong offensive security skills. Prediction, medium confidence: Carbon reaches developers within weeks, at the intro price. The part nobody outside Google can check yet is how much of that 88.2% survives once the guardrails go back on.