Gemini 4 Carbon: what it is and who can use it

Gemini 4 Carbon is Google's new top model, announced October 28, two weeks after Gemini 4 Argon, and it is built for cybersecurity work and long, multi-step agent jobs. You probably can't use it yet: for now it goes only to a vetted group of cyber defenders in Google's Fairwind Program, and developers, businesses and the Gemini app get it later, with no date set.

Google's accounts posted within about two minutes of each other. @GoogleDeepMind called it "our next frontier model." @sundarpichai led with the ranking: first place on the Artificial Analysis Intelligence Index v5.0, at 59. And @OfficialLoganK said it's "a big jump over Argon which we released just 2 weeks ago."

Our Gemini 4 Argon launch post went up four weeks ago, so that's fast.

What Gemini 4 Carbon can do

According to Google's announcement, Carbon is its "strongest model with state-of-the-art cybersecurity and long horizon capabilities." In practice that covers agents that work on hard engineering tasks for days at a stretch, or trace a bug across a big codebase back to its cause.

Google covered the security pitch in three steps: finding vulnerabilities, proving they can be exploited, and patching them at the root. Google says it ran the outside security benchmarks without guardrails switched on. It also says Chrome's security team got patches more than 10 times better than those from an unnamed rival lab's frontier model.

Here's how Carbon compares with Argon, using figures from Google's post and VentureBeat's coverage:

Carbon doesn't win everything. @ArtificialAnlys puts it first overall but second to Claude Opus 5.5 on its Coding Agent Index. VentureBeat lists that result as 62 for Claude Code with Opus 5.5 against 59 for Carbon. On VentureBeat's table, Opus also leads the OSWorld computer-use test at 87.2%.

Watch the hallucination numbers too. Per VentureBeat, Carbon gets more answers right on Artificial Analysis's knowledge test, but its hallucination rate there rose from 4% to 11%. Google's internal testing says the opposite: about half as many unsupported claims as Argon.

Who can use Gemini 4 Carbon, and what it costs

Right now, only a subset of cyber defenders can use it. Google says it will widen access through Fairwind before a general release, the same staged path Argon took. Argon launched to defenders on September 30 and reached paying users on October 13, which we covered in is Gemini 4 Argon free. Two weeks isn't a promise, though. Google hasn't given Carbon a date.

Google has published the price already:

Cached input is 95% off. Those rates match Argon's. Google hasn't said when the introductory period ends, and it hasn't published a context window or an API model name.

How to get access now

If your organization does defensive security work, the route is the Fairwind Program page, through its "Apply for access" form. The page still describes Argon as the program's model and doesn't mention Carbon yet. Its rules include:

• Priority goes to governments, critical infrastructure operators (healthcare, telecom, energy, finance), core tech platforms and academic labs working on defensive benchmarks.
• Use is limited to authorized threat simulation, reverse engineering and malware analysis.
• You need user-level authentication, phishing-resistant MFA and tracking of each employee's use.
• Only internal security, incident response or penetration testing teams get access, and you can't resell or share it.

For everyone else, the practical move is to build on Argon now. It's in the Gemini API and the paid Gemini app tiers, and since Carbon has the same price, swapping models later should mostly be a configuration change.

My read

Google now ships in small, fast steps, and security teams go first each time. Given what these models can do to software, I think that's right. This morning we wrote about the exploit chains an unreleased Claude model built, and the labs seem to agree that defenders should get a few weeks' head start.

The rankings are also close. Carbon, Opus 5.5 and GPT-6 Astra sit within two points of each other on the Artificial Analysis index, so the right model for your work depends more on the task than the leaderboard.

What this means for businesses adopting AI

The top model changes every few weeks now, and the price can double when a promotion ends. Workflows that hold up are the ones where the model is a single setting you can switch without rebuilding anything. If you'd like help figuring out which of your processes are worth running on a frontier model, New Face Design offers a free process audit.