Google built the smartest AI on the charts, and only hackers get it

Google has unveiled Gemini 4 Carbon, its new flagship model, just two weeks after Gemini 4 Argon came out.

Carbon now sits at number one on the Artificial Analysis Intelligence Index v5.0, an independent combined benchmark of AI models, with a score of 59. The focus is cybersecurity and long, unsupervised work. It scores 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71% on CWE-bench for patching vulnerable code. On WebDev Arena it climbed to fourth place.

The most striking part of Google's post: internally, Carbon has been left running for days at a time on hard engineering problems, pushing through dead ends to root-cause and fix subtle bugs in large codebases. The Chrome team says it produced high quality patches more than 10 times better than another lab's large frontier model (Google doesn't say which). On internal tests it writes secure code more than twice as often as Argon.

Regular users can't have it yet. Cyber defenders get it first through Google's Fairwind program, and Google is taking part in the US government's voluntary pre-release model testing. Developers, businesses and consumers are promised access "as soon as possible". API pricing at launch matches Argon: $2 per million input tokens and $10 per million output tokens.

Bottom line: Google is handing its strongest AI to the defenders first, because it is that good at finding holes.