Google has a new strongest model, and it is not really for you — at least not yet. Gemini 4 Carbon, announced Wednesday by Google DeepMind, tops the updated Artificial Analysis Intelligence Index at 59 and is pitched squarely at cybersecurity: finding vulnerabilities, proving they can be exploited, and patching them.

The numbers Google leads with are striking, if you take them on Google's terms. Carbon scores 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Internally, Google says Carbon generates secure code at more than twice the success rate of Argon.

That two-week cadence is the part worth pausing on. Argon launched two weeks ago; Carbon already succeeds it. The frontier is no longer moving in quarterly leaps. It is moving in fortnights, and each new model arrives with a benchmark sheet that makes the last one look like a draft. The question for everyone outside the labs is no longer whether the next model is better — it almost certainly will be — but whether the gains land anywhere you can actually use them.

For Carbon, the answer is: not yet, and deliberately so. Google is rolling it out first to a subset of cyber defenders through its Fairwind Program, and says it is engaged in the U.S. government's voluntary pre-release testing process. Developers, businesses, and consumers get it "as soon as possible" — no date attached. The caution is understandable for a model whose whole selling point is that it is unusually good at breaking into software.

There is a second, quieter claim that may matter more in the long run. Google says Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, cut unsupported claims roughly in half versus Argon, and resisted every evaluated prompt-injection attack on Gray Swan's benchmark. If those numbers hold up outside Google's labs, that is the kind of improvement that makes agentic coding workflows actually trustworthy, rather than just faster.

The price is already set: $2 per million input tokens and $10 per million output tokens at launch, matching Argon's introductory rates, with cached input at 95% off. So the model exists, the price exists, the benchmarks exist — only the access does not. Watch the Fairwind rollout. How quickly Carbon moves from defenders to developers will tell you how confident Google really is in its own safeguards.

— via Google Blog