Google's Gemini 4 Carbon Tops a Major AI Leaderboard, But You Can't Use It Yet
Google announced Gemini 4 Carbon, its strongest AI model yet, with record cybersecurity scores. Only a small group of vetted cyber defenders get access first.
88.2%. That’s the score Google says its newest AI model hit on SEC-bench Pro, a test of finding real security flaws buried in real-world software. The model is called Gemini 4 Carbon, Google announced it on October 28, and it was one of the most talked-about tech stories of the day on X and Hacker News.
There’s one big asterisk, though: almost nobody can use it yet.
What Google announced
Google calls Carbon its “strongest model” so far, built for cybersecurity and for long, multi-step tasks. According to Google, it ranks #1 on the newly updated version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The Verge notes that puts it ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra on that leaderboard.
It also arrives just two weeks after Google launched Gemini 4 Argon. That’s a fast turnaround, even by AI-industry standards.
Built to keep going for days
The headline feature beyond the benchmarks is stamina. Google says that inside the company, Carbon has run for days on hard engineering problems, pushing through dead ends to do things like root-cause and patch subtle bugs across big codebases.
Google also published a stack of security numbers:
- SEC-bench Pro (finding vulnerabilities): 88.2%
- ExploitGym (checking whether a flaw can really be exploited): 54.8%
- CWE-bench (writing the fix): 71.0%
Google adds that Carbon cut unsupported claims roughly in half compared with Argon in its internal testing, and blocked every attack in Gray Swan’s prompt-injection benchmark. Keep in mind these are Google’s own reported results, so independent testing will matter.
Why you can’t try it
Google is deliberately rolling Carbon out slowly. Starting this week, only a subset of cyber defenders get access, and Google will gradually widen that through its Fairwind Program, which is aimed at groups like governments, healthcare providers and telecom companies. Google says it’s also taking part in the U.S. government’s voluntary pre-release testing process for AI models.
Developers, businesses and regular users come later, “as soon as possible” in Google’s words, with no date given. When it does open up, Google says it will start at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens.
Why it matters
An AI that’s great at finding security holes is a double-edged sword. In defenders’ hands it can find and patch bugs before attackers do. In the wrong hands, those same skills could help someone break in. Google’s slow, vetted rollout is its answer to that problem, and it’s a sign of how seriously AI companies are starting to treat cyber capabilities.
Would you trust an AI to hunt for bugs in your code, or does a model this good at security make you a little nervous?
Sources: Google blog, The Verge, Google DeepMind Fairwind Program
