Google Unveils Gemini 4 Carbon, and Security Teams Get It First

Google announced Gemini 4 Carbon on Wednesday and calls it its strongest model yet. It comes just two weeks after Gemini 4 Argon arrived in the Gemini app.

You can't use it yet. For now, only a small group of cybersecurity teams has access, since Carbon is especially good at finding and fixing security flaws in software.

Carbon is built for jobs that take days

Google built Carbon for long, multi-step work, the kind where a model has to keep going through dead ends without someone steering every step.

Google says its own teams have watched Carbon run for days on hard engineering jobs, like tracking down the root cause of a subtle bug in a huge app's code and then fixing it. For you, that would mean handing it a bigger project in one go and checking in on it less.

Google also says it got better at building websites and web apps, and at understanding images and other files alongside text.

Why security teams get Carbon before you do

A model that's good at finding flaws in software can help the people who fix those flaws, and in the wrong hands it could also help the people who exploit them. Google says Carbon covers the whole defender's job: finding a weakness, checking whether it can actually be used in an attack, and writing the fix.

Google's own Chrome team reported that Carbon wrote much better fixes for security bugs than another lab's top model.

So Google is starting small. According to Google DeepMind's post, the rollout goes like this:

1. A small group of security teams gets Carbon this week.
2. More trusted partners come in gradually through Google's Fairwind Program.
3. Developers, businesses and regular users get it "as soon as possible," once Google has gathered feedback and tuned its guardrails.

Google is also part of the US government's voluntary program for testing new models before they come out, one more check before Carbon goes wide.

A model that tells you what it couldn't check

Google says Carbon makes about half as many claims it can't back up as Argon. When it hits a gap, it's meant to say so in the work it hands you, or list what's still missing, instead of making something up to fill the hole.

It's also meant to ask when a request is unclear. Google says Carbon pulls up the relevant sources and points out where they disagree before it acts, instead of guessing and pushing ahead.

And it's meant to do what you asked without wandering off. Google says Carbon is less likely to go beyond what you asked for, and less likely to quietly do less than you asked, on long multi-step plans.

When Carbon could reach the Gemini app

There's no date for everyone else. Argon gives a rough idea of the pace: it went to security teams first on October 1, then landed in the Gemini app for Google AI Pro ($19.99 a month) and Ultra (from $99.99) subscribers on October 13.

Google hasn't promised the same pace this time, and Carbon's rollout has an extra step for trusted partners. It also hasn't said which Gemini plans will get Carbon or whether it will cost more. For developers, Google says it will cost the same as Argon.

Why I'm curious after Argon let me down

I was very underwhelmed with Gemini 4 Argon in my test. It came last on all four builds I gave it against the top models, and on the $99 a month plan I hit the daily limit after about one build.

Releasing another model right after it is an interesting move, and honestly the last thing I expected from Google. It's good to see they have more in the pipeline.

Like Carbon, Argon went to security teams first and landed in the Gemini app less than two weeks later, so Carbon could show up soon. I haven't been able to try it, since nobody outside those security teams can yet.

Here's my Argon test:

Sources

• Google: Gemini 4 Carbon, our next frontier model
• Google DeepMind: Gemini 4 Carbon, our next frontier model
• Google DeepMind: Fairwind Program
• Google: Introducing Gemini 4 Argon
• @sundarpichai on X: introducing Gemini 4 Carbon