Google announces its top model
Google on Oct. 28 (early Oct. 29 Japan time) announced "Gemini 4 Carbon," its new top-end AI model. It succeeds Gemini 4 Argon, whose general availability was expanded on Oct. 13, an update coming just two weeks later. According to Google, it scored 59 on the Artificial Analysis Intelligence Index, widely referenced as a measure for comparing the overall intelligence of AI models, taking first place.
However, the first to get access are only cybersecurity defense organizations vetted by Google, and no timeline has been given for availability to developers or general users.
Logan Kilpatrick, who leads AI Studio at Google, also emphasized the size of the gain over Argon.
Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago. Gemini 4 Carbon is SOTA on cyber, long horizon tasks, ranks #1 on Artificial Analysis V5, significantly improves on usability in domains like coding, and is more aligned.
There were signs before the announcement. On Oct. 19, U.S. outlet The Information reported that Gemini 4 Carbon would arrive this month, and on Oct. 21, testing of a model believed to be Carbon was reported on LMArena, a site that compares AI models anonymously.
Top scores in finding and exploiting vulnerabilities
Carbon made especially large gains in the ability to find software flaws (vulnerabilities), actually exploit them and patch them. In evaluations published by Google, it set the highest scores at the time of publication on two benchmarks: SEC-bench Pro for discovery and ExploitGym for exploitation.
According to Google, when the Chrome development team had it create fix patches, the results were more than 10 times better than those of other companies' large models. In internal evaluations, the rate at which it writes secure code was more than double that of the previous Argon.
Why defenders get it before the public
The same capabilities serve as a repair tool for defenders and an attack tool for attackers. Google therefore chose an order that gives it to defenders first, creating time to patch flaws in existing software. The point of access is the "Fairwind Program," a vetted program in which more than 650 organizations participate. Google also took part in voluntary safety testing of AI models before release conducted by the U.S. government.
Anthropic, with Glasswing, and OpenAI, with Daybreak, set up programs with the same purpose earlier and offer models in advance only to vetted companies. That means all three major U.S. companies have taken the same steps.
On the other hand, some point out that high-performance models released openly by China's MiniMax, Alibaba and Z.ai are available to anyone, and that attackers are already outside these restrictions.
A 1-point margin at the top
Although it leads the overall evaluation, its margin over the rest is small. On Artificial Analysis's index, Anthropic's Claude Opus 5.5 scored 58 and OpenAI's GPT-6 Astra 57, 1 to 2 points behind Carbon.
Google has previously taken first place with Gemini 3 Pro in November 2025 and Gemini 3.1 Pro in February 2026, but in both cases rivals caught up within weeks. In a separate evaluation focused on writing code, Anthropic's Claude Code scored 62, above Carbon's 59.
The title of No. 1 does not last long, and rankings also change by use. For companies considering adoption and investors following competition among AI companies, looking at which model is strongest for their own use is less likely to lead to misjudgment than the ranking by overall score.
Pricing and availability
The price developers pay to build it into their own services (API pricing) has been kept at the same level as Argon for the time being. Per 1 million tokens, the units of text the AI reads and writes, input is $2 and output is $10. However, this is an introductory price, and after the period ends it doubles to $4 for input and $20 for output. The length of the introductory period has not been disclosed.
Availability to developers, businesses and general users is described as coming as soon as possible, with no specific date given. Google says it will expand access after adjusting safety measures based on feedback from defenders.
