Gemini 4 Carbon: Google's new model for cybersecurity

On October 28, Google introduced Gemini 4 Carbon, its new frontier model. It was announced by Koray Kavukcuoglu (Google DeepMind, Chief AI Architect at Google), who described it as a model with “state-of-the-art cybersecurity and long horizon capabilities”. For now, only selected companies in the cybersecurity industry will get access.

What’s new about it in the AI landscape?

Carbon arrives just two weeks after Gemini 4 Argon, which debuted on October 13. Google is focusing on three things:

1. Long agentic tasks. According to the announcement, Google teams have seen Carbon run for several days in a row, working its way past dead ends in multi-step tasks, such as finding the root cause of bugs and patching them in large codebases and low-level systems.
2. Cybersecurity. The model is meant to find vulnerabilities, confirm them and write fixes. According to Google, in internal tests it generates secure code more than twice as effectively as Argon.
3. Web development. On WebDev Arena it gained 89 Elo points and sits in 4th place with a score of 1769.

On top of that, according to Google, the cloud security platform Wiz sees Carbon leading its web application penetration testing benchmarks, which supports Wiz’s Scan for Good initiative.

How does Gemini 4 Carbon score?

On the Artificial Analysis Intelligence Index v5.0, Carbon has 59 points and first place. For security benchmarks, Google gives these numbers:

Benchmark	What it measures	Score
SEC-bench Pro	finding vulnerabilities	88.2%
ExploitGym	writing exploits	54.8%
CWE-bench	patching	71.0%

Interestingly, Google notes that it ran the external benchmarks without guardrails, so they measure the model’s raw capabilities.

Is Carbon safer than Argon?

According to Google, yes. On the Gray Swan benchmark for indirect prompt injection attacks, the attack success rate dropped to 0.0% (Argon had 0.7%). In Terminal-Bench 4.0, successful task cheating dropped to 0.0%, and cheating attempts to 0.4%.

The most interesting number, though, concerns agents working in real code: alerts about serious misalignment in programming environments dropped by almost 80%. In addition, according to internal tests, the model makes unsupported claims less often (roughly half as often as Argon), is 26% less likely to do more than it was asked, and 29% less likely to do less. The details are in the Gemini 4 safety report.

How can you try Gemini 4 Carbon?

As of today, you most likely can’t. Google is rolling it out in stages:

1. Starting this week: a selected group of cybersecurity defenders.
2. Next: more trusted partners through the Fairwind program, which you can apply to via a form. Priority goes to governments, critical infrastructure operators (hospitals, telecoms, energy, finance) and large tech platforms, and Google vets the candidates.
3. “Soon”: developers, businesses and regular users, once Google has gathered feedback and refined the safeguards.

The launch price is the same as Argon’s: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached tokens. After the promotional period it will be $4 and $20. Google hasn’t said when that period ends for Carbon.

Google also takes part in voluntary pre-release model testing run by the US government.

What should you pay attention to?

First, who gets the model first. Carbon goes to defenders first, just as Claude Mythos went to Bank of America earlier. Models that find and exploit vulnerabilities keep getting stronger, and the labs would rather defenders have them before attackers do. The PHP2Shell vulnerabilities in PHP, found partly with Argon’s help, showed this isn’t just theory.

Second, where the numbers come from. Most of the results come from Google or from internal tests, and they can only be checked independently once the model reaches a wider audience.

Third, the pace. Two weeks between Argon and Carbon is a very fast cycle, even by today’s AI standards.

FAQ

What is Gemini 4 Carbon?

It’s a new frontier model from Google DeepMind, announced on October 28, 2026. It’s built for cybersecurity and long agentic tasks, which according to Google can last as long as several days.

When will Gemini 4 Carbon be available to everyone?

Google hasn’t given a date. Selected defenders and Fairwind program partners get it first; developers, businesses and users “soon”.

How much does Gemini 4 Carbon cost?

At launch, $2 per million input tokens and $10 per million output tokens, the same as Gemini 4 Argon. After the promotional period, $4 and $20.

How is Gemini 4 Carbon different from Gemini 4 Argon?

According to Google, Carbon does better at cybersecurity, multi-day agentic tasks and web development, and in safety tests it is harder to manipulate. Argon is already available through the API and in the Gemini app; Carbon isn’t yet.