Google’s Gemini 4 Carbon Tops the AI Rankings, but Only Cyber Defenders Can Use It for Now

Google announced a new flagship AI model on October 28, 2026, and then did something unusual with it. Instead of opening it to developers and paying customers right away, the company handed it first to a small group of cyber defenders. The model is called Gemini 4 Carbon, and according to Google it is the strongest model the company has built so far. The catch is that most people cannot try it yet, and Google has not said when that will change.

That staged release is the real story here. Carbon sits at the top of a widely watched industry ranking, it carries a published price, and it follows a previous model that reached customers only two weeks earlier. Yet Google is treating it like a tool that needs vetting before it spreads.

What Google announced

Gemini 4 Carbon is Google DeepMind’s newest frontier model, built for cybersecurity work and long software engineering jobs. Google describes it as a model that can find software flaws, test whether an attacker could exploit them, and repair the root cause rather than just patching the symptom.

The timing is striking. Carbon arrived just two weeks after Gemini 4 Argon, its predecessor, became broadly available to developers and businesses on October 13. Google says Carbon improves on Argon on every benchmark it published. It also returns Google to the top of the Artificial Analysis Intelligence Index. Carbon scored 59 on version 5.0 of the index, ahead of Anthropic’s Claude Opus 5.5 at 58 and OpenAI’s GPT-6 Astra at 57.

The security scores Google is leading with

Google framed the launch around cybersecurity rather than general ability, and the numbers it chose make that clear. On SEC-bench Pro, a test of vulnerability discovery, Carbon scored 88.2 percent. On ExploitGym, which checks whether a model can show that a flaw is actually usable in an attack, it scored 54.8 percent. On CWE-bench, a test of writing patches that fix a flaw while keeping the software working, it scored 71.0 percent.

Those figures look stronger when placed next to rivals. According to Google’s benchmark table, GPT-6 Astra scored 42.4 percent on ExploitGym and Claude Opus 5.5 scored 34.5 percent. Google also reports that its Chrome team found Carbon generated high quality patches more than ten times better than another lab’s large frontier model in its own tests, though it did not name the rival model or publish the method behind that comparison. On a separate test of indirect prompt injection attacks, Google reports a 0.0 percent attack success rate for Carbon. None of these headline numbers have been reproduced outside Google so far, so treat them as Google’s claims for now, not independent verification.

Why defenders get it first

Carbon’s first users are a subset of cyber defenders reached through Google’s Fairwind Program. Fairwind is an early access program for high priority defensive work, covering governments, national cyber authorities, and critical infrastructure operators in areas such as health care, energy, telecom, and finance. Reports put the program at more than 650 partners. Partners go through background checks and may use the model only for defensive tasks such as authorized threat simulation, reverse engineering, and malware analysis.

The logic is straightforward. A model that is good at finding and exploiting flaws is useful to defenders, but the same capability in the wrong hands is a risk. Google says it is taking part in the United States government’s voluntary pre-release model testing process and plans to keep adjusting Carbon’s safeguards before widening access. This is also not a brand new playbook. Google took a similar path with Argon, opening it first to cybersecurity partners before a wider release. What is different this time is the gap. Argon reached paying developers within about two weeks, while Carbon has no published date for developers, businesses, or consumers. Google has said only that broader access will come as soon as possible.

What it will cost when it does arrive

Even without a release date, Google has published Carbon’s price. It launches at the same introductory rates as Argon: 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input priced 95 percent lower. After an introductory period that Google has not dated, the rates rise to 4 dollars per million input tokens and 20 dollars per million output tokens.

That pricing tells its own story. Google is positioning Carbon as a premium tool for serious engineering and defense work, not as a cheap everyday chatbot.

Why the top spot may not last

Google’s return to the top of the rankings is real, but history suggests it may be brief. The company’s last two turns at the top lasted only weeks before OpenAI and Anthropic caught up, and both rivals are widely expected to ship new frontier models soon. Carbon leads its rivals by a single point on the index, and it does not lead in every category: reports note that it led four of the index’s nine evaluations but not coding or knowledge work.

For now, Carbon is a model to watch rather than a model to try. Security teams inside the Fairwind Program will learn first whether it lives up to its scores in real investigations. Everyone else gets a ranking, a price list, and a wait. If Google’s past pattern holds, that wait could be weeks. If the safeguards need more work, it could be longer. Either way, the launch marks a shift: the frontier is no longer defined only by who builds the strongest model, but by who gets to use it first.

Frequently Asked Questions

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google DeepMind’s newest flagship AI model, announced on October 28, 2026. Google says it is built for cybersecurity work, such as finding and patching software flaws, and for long software engineering tasks.

Why can most people not use Gemini 4 Carbon yet?

Google is releasing Carbon first to vetted cyber defenders through its Fairwind Program while it continues testing safeguards. Developers, businesses, and consumers will follow, but Google has not announced a date.

How much will Gemini 4 Carbon cost?

Google lists introductory prices of 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input 95 percent cheaper. After an unspecified introductory period, the rates rise to 4 dollars and 20 dollars per million tokens.