Google unveils Gemini 4 Carbon, rolls out to vetted cyber defenders

Google on Wednesday announced Gemini 4 Carbon, a new flagship artificial intelligence model that it says is built for cybersecurity work and long-horizon tasks, but the company is rolling it out first to a limited group of vetted defenders rather than releasing it broadly.

In a blog post titled “Gemini 4 Carbon: our next frontier model,” Koray Kavukcuoglu, Google’s senior vice president at Google DeepMind and chief AI architect, described Carbon as Google’s strongest model for “state-of-the-art cybersecurity and long horizon capabilities.” Google said the model is designed for sustained agentic workflows — meaning AI systems that can carry out multistep work over longer periods — and harder engineering tasks. The company also said Carbon ranks No. 1 on Artificial Analysis Intelligence Index v5.0 with a score of 59 and posted strong results on several cyber benchmarks. But those specific Carbon entries were not yet visible on public benchmark pages when checked, meaning the scores now come from Google’s announcement rather than independently published leaderboards.

The rollout is the bigger immediate story. Google said “starting this week” it will make Carbon available to “a subset of cyber defenders” and then expand access to additional trusted partners through its Fairwind Program, a Google DeepMind initiative that gives vetted cyber defenders early access to frontier models. Google said it will use that phased release to gather feedback and refine guardrails before making Carbon available more broadly to developers, businesses and consumers. It did not provide a general-availability date.

Google said Carbon builds on Gemini 4 Argon, which it said launched two weeks earlier. In its announcement, the company reported Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench (Collinear), benchmarks that test tasks such as finding software vulnerabilities, checking whether they can actually be exploited and helping patch flaws. Google also reported more than a 2x uplift over Argon for generating secure code, said the Chrome team found Carbon performed more than 10x better than another lab’s large frontier model at generating high-quality patches, and said Wiz observed Carbon leading on web penetration-testing benchmarks. Those results were reported by Google and were not independently confirmed in public leaderboard postings at the time checked.

Google also disclosed pricing. It said the introductory price will be $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95% from the input token price. After the introductory period, Google said pricing will rise to $4 per million input tokens and $20 per million output tokens.

The release continues Google’s fast-paced Gemini 4 rollout. Gemini 4 Argon was announced in mid-October, and Carbon follows about two weeks later with a much sharper emphasis on cybersecurity. That focus helps explain the restricted launch. The benchmarks Google highlighted measure capabilities that can help defenders find, verify and fix weaknesses, but they are also dual-use tools that could be misused if made widely available without controls.

Google said it is also participating in the U.S. government’s voluntary pre-release model testing process. That fits into a broader environment of voluntary oversight after NIST, the National Institute of Standards and Technology, the National Security Agency and the Cybersecurity and Infrastructure Security Agency published guidance on AI incident readiness and response earlier this month. For now, though, Carbon’s broader release will depend on what Google learns from its defender-first rollout and further guardrail work.