Google introduces Gemini 4 Carbon, a powerful new language model that focuses specifically on cybersecurity and complex engineering tasks. According to Google, the model should push the boundaries of AI agents and software security.
Google has come up with yet another new model. The announcement follows just two weeks after the official launch of Gemini 4 Argon, unveiled in early October. For the naming, the engineers once again threw a dart at the periodic table of elements. The name Gemini 4 Carbon could also be a reference to the growing emissions in data centres that come with the rapid succession of model launches.
Google presents Carbon as its strongest model to date. Two weeks ago, it said the same about Argon. The model claims first place on several benchmarks. With it, Google wants to speed up the pace of technological progress while strengthening safety measures at the same time.
Focus on cybersecurity and coding
Gemini 4 Carbon stands out for its performance in cybersecurity defence. In the SEC-bench Pro, the model scores 88.2 percent at detecting vulnerabilities in software. It also scores high on fixing bugs in the CWE-bench, with 71.0 percent.
For validating risks through exploitation, Carbon achieves 54.8 percent on the ExploitGym benchmark. Internal tests at Google show that the model generates secure code twice as often as its predecessor, Argon. The Chrome team even reported that Carbon performed ten times better than other models at creating patches.
Besides security, the model is designed for long-running tasks in which AI agents have to keep working for days. This is essential for complex engineering problems, such as tracking down subtle bugs in huge codebases or low-level systems.
Safety and rollout
Google is taking a cautious approach to the availability of the model. This week, Carbon will first be made available to a select group of cyber defenders. Through the Fairwind Programma, more trusted partners will get access to the technology later. Google did the same for Argon, which is now more widely available.
The model’s safety has been improved by training on complex tasks and reducing erroneous claims. Carbon halves the number of unsubstantiated statements compared to Argon. Resistance to prompt injection has also improved: in tests, the model achieved a perfect score.
The price of using Carbon remains the same as that of Argon. Users pay 2 dollars per million input tokens and 10 dollars per million output tokens. Google is currently working on the further rollout to developers, businesses and consumers through its regular product surfaces.
Pause over
It is remarkable that Google is launching two frontier models in little more than a month. Not so long ago, the big AI companies seemed to have come to their senses and vowed to take it easier. A “pause” was needed to get security in order. In the end, that produced little more than a sheet of A4 signed at the White House, and in the meantime the race is carrying on as before.
Moreover, practice shows that businesses do not need to have the biggest, most expensive and most powerful model. Many tasks for which businesses use AI can be done more cheaply and efficiently with smaller LLMs. Google seemed to have understood that message by betting on the Gemini Flash models, for which it also launched new models faster than it changes its underwear, but with first Argon and now Carbon it is going for bigger, bigger, biggest again.