Google releases Gemini 4 Carbon, claims top AI benchmark and cybersecurity gains

Google released a new frontier model called Gemini 4 Carbon on Wednesday, saying it delivers “significant gains” in coding, agentic tasks, and cybersecurity, according to the company’s announcement. Google says Carbon now holds the top spot on the Artificial Analysis Intelligence Index, version 5.0, a third-party ranking of overall model capability, ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra.

Carbon is an upgrade to Argon, the model Google released earlier this month. For now, access is limited to a select group of cybersecurity defense companies. Google says it will expand availability to other partners in its Fairwind Program over time, though it has not given a timeline for that wider rollout.

A cybersecurity push

Google is framing Carbon’s cybersecurity ability as its headline feature, describing a “significant leap” in cyberdefense that makes the model better at finding vulnerabilities and writing secure code. The company points to two benchmark results to back that up: Carbon holds the number one spot on SEC-bench Pro, which measures how well a model can discover real-world software bugs, and on ExploitGym, which measures a model’s ability to turn a known vulnerability into a working exploit. Google says Carbon beat GPT-6 Astra on the SEC-bench Pro benchmark specifically, though it did not publish the full scores for either benchmark in its announcement.

Google also says Carbon is built to work on longer, more complex jobs. Internal teams have reportedly seen the model “run for days” to work through dead ends on multi-step engineering tasks, according to the company. Google adds that Carbon is better at avoiding shortcuts, resisting prompt injection attacks — a technique where hidden instructions in text or data trick an AI into acting against its user — and collaborating with the people using it.

On honesty, Google says Carbon cuts “unsupported claims” by half compared with Argon, though the company did not describe how it measured that figure or what counts as an unsupported claim. As with the benchmark rankings, that comparison comes from Google itself and has not been independently verified.

Pricing stays the same as Argon’s introductory rate: $2 per million input tokens and $10 per million output tokens. Input tokens are the chunks of text a model reads when given a prompt; output tokens are the chunks it generates in response. Keeping the price flat suggests Google is trying to make the upgrade an easy switch for existing Argon customers rather than a premium new tier.

What happens next depends on how Carbon performs once it reaches more than a handful of cybersecurity partners. Independent researchers have not yet had a chance to test Google’s benchmark claims themselves, and the model’s broader rollout through the Fairwind Program will be the first real test of whether its coding and security gains hold up outside Google’s own evaluations.