Google's Gemini 4 Carbon: what it does and who can use it

Google DeepMind has announced Gemini 4 Carbon, a new AI model it calls its strongest yet. It was unveiled on 28 October 2026, two weeks after the company launched Gemini 4 Argon.

Carbon is aimed mainly at writing and fixing software and at helping protect computer systems from attack. Below is what Google says it can do, how it has tried to make it safer, and what is still unknown.

What Google has announced

Google describes Carbon as a 'frontier model', meaning one of the most advanced AI models a company has built, at the edge of what the technology can currently do.

The Artificial Analysis Intelligence Index compares AI models across a range of tests, and Google says Carbon now ranks first there with a score of 59. On WebDev Arena, a comparison focused on building websites, it climbed to fourth place.

Google puts a lot of weight on what it calls 'long-horizon' work: tasks that take many steps and a long time to finish. Google says its own teams have watched Carbon keep working for days on hard engineering problems, pushing past dead ends to track down and fix subtle bugs hidden in very large amounts of code. It is closer to handing a colleague a whole project than asking them a quick question.

What it is good at

The main focus is cybersecurity, keeping computer systems safe from attackers. Google tested Carbon on three outside tests, often called benchmarks, which work a bit like standard exams for AI models:

• Finding weaknesses: 88.2% on SEC-bench Pro, which checks how well a model uncovers flaws in real software.
• Checking how serious a weakness is: 54.8% on ExploitGym, which tests whether a model can show that an attacker could actually use a flaw. Defenders use this to decide what to fix first.
• Fixing the problem: 71.0% on CWE-bench, which tests whether a model can turn a discovered flaw into working code that repairs it.

Google calls these leading results. It also reports that on its own internal tests Carbon was more than twice as successful as Argon at writing secure code, and Google's Chrome team found it did more than ten times better than another lab's large model at producing high-quality fixes.

Google adds that Carbon has improved at science tasks, at building complete websites, and at understanding more than just text.

How Google says it has made it safer

A model that can find and exploit weaknesses could be dangerous in the wrong hands, so Google spends much of its announcement on safety.

One concern is models taking shortcuts, such as gaming a test instead of solving it. Google says Carbon cut successful cheating on a coding test called Terminal-Bench 4.0 to zero, and attempts to cheat fell to 0.4%.

Another is 'prompt injection', where someone hides instructions inside a document or web page in the hope that the AI will read and obey them. Google says Carbon resisted every attack in a test of this from Gray Swan.

According to Google, Carbon also makes roughly half as many unsupported claims as Argon. It flags gaps in its evidence and asks the user when something is unclear. It goes beyond what it was asked to do 26% less often and falls short of what it was asked 29% less often.

Can you use it yet, and what will it cost?

For most people, not yet. Google is rolling Carbon out slowly. From this week a small group of cyber defenders can use it, and more trusted partners will be added gradually through what Google calls its Fairwind Program. Google is also taking part in the US government's voluntary process for testing models before release.

Developers, businesses and everyday users will get access later, once Google has gathered feedback and improved its guardrails, the built-in limits that stop misuse. Google says wider access will come as soon as possible but gives no date.

Models like this are usually charged by the 'token', a small chunk of text the model reads or writes. Carbon will start at an introductory price of $2 per million input tokens and $10 per million output tokens, the same as Argon. After that the price rises to $4 and $20. Google has not said how long the introductory period lasts.

What we still don't know

Most of these figures come from Google itself, and several come from internal tests that outsiders cannot check.

The cybersecurity scores were measured with guardrails switched off, so the version people eventually use may behave differently.

There is also no firm date for when small businesses or the public can try Carbon.

Source: Gemini 4 Carbon, Google DeepMind. Listed in AI News Week 45, where we track every AI release each week.