NederlandsGoogle presents Gemini 4 Carbon: better at finding and fixing security flaws
Two weeks after Gemini 4 Argon, Google DeepMind has released Carbon, which Google calls its strongest AI model so far. It excels at computer security and long tasks. Security experts get access first.
Google DeepMind has announced Gemini 4 Carbon, a new AI model at the top of the Gemini family. According to Google, it is its strongest model so far. It is especially good at computer security and at long, complex tasks that it carries out on its own. Carbon builds on Gemini 4 Argon, which came out only two weeks earlier.
Finding, exploiting and fixing security flaws
Software companies are constantly looking for weak spots in their programs before attackers find them. That work is slow and takes a lot of expertise. Google says Carbon helps a great deal with it.
According to Google, Carbon achieves the highest scores on three public tests for security work. It found 88.2 per cent of the flaws in a test called SEC-bench Pro. On ExploitGym, a test that measures whether a model can actually take advantage of a flaw, it scored 54.8 per cent. And on CWE-bench, a test by the company Collinear for repairing flaws, it reached 71.0 per cent. Google tested the model without its built-in safety brakes, to see what it can really do.
Google also sees progress inside the company itself. According to Google, Carbon produces secure code twice as often as Argon. The team behind the Chrome browser reports that Carbon's fixes for errors were more than ten times better than those of a large AI model from another company. According to Google, internal teams saw the model keep working for days on difficult jobs, such as tracking down and fixing hidden errors in large programs.
More reliable behaviour
Google has also paid attention to how honestly the model works. In a programming test in which a model can try to cheat, Carbon never succeeded in cheating, according to Google. The model makes claims it cannot back up about half as often as Argon.
Carbon is also better at resisting hidden instructions in texts and web pages, which attackers use to try to mislead an AI model. In a test by the security company Gray Swan, such an attack still worked on Argon in 0.7 per cent of cases. With Carbon, it no longer happened.
According to Google, Carbon is at the top of a general ranking by the independent organisation Artificial Analysis, with a score of 59.
Security experts first
Google is releasing the model step by step. In the last week of October 2026, a group of security experts gets access. More trusted partners follow after that, through a programme called Fairwind. Developers, businesses and ordinary users come later. Google also has the model tested in advance in a voluntary programme run by the US government.
For developers, Carbon initially costs the same as Argon: 2 dollars per million tokens (small pieces of text) that the model reads, and 10 dollars per million tokens that it writes. After this introductory period, that becomes 4 and 20 dollars.
First impressions from testers
Testers who got their hands on the model earlier are sharing their first impressions, the tech site TestingCatalog reports. According to them, Carbon is clearly better than Argon at 3D work. It is, however, two to four times slower, and it is not yet good at designing websites and apps. These impressions come from individual comparisons and tasks the testers picked themselves, not from fixed tests. It is also not certain which version of the model the testers received.
Earlier in October 2026, Business Insider reported that Google was already using Carbon internally in its own programming tool. An employee there compared it with Opus 5.5, Anthropic's strongest programming model, but also said it still needed more testing.
The performance figures come from Google itself. Independent measurements will follow as soon as more people can use the model.
What does this mean for the road to the singularity?
From Argon to Carbon in two weeks: new AI models are following each other ever faster. A model that can keep working on its own for days on errors in large programs is taking over more and more real technical work. That work is also needed to build new AI models. Vedant Misra of Google DeepMind responded to the news about Carbon with a reference to "recursive self improvement": AI that helps make better AI. That Google gives the model to defenders first shows how stronger AI can also make the internet safer. Now that the model is becoming more widely available, it will become clear how well it works in practice.
What is the technological singularity? Read the explanation →
