New AI model released by Google

Sundar Pichai, CEO of Google, confirmed on X on Wednesday, October 28, that new AI models would be released at regular intervals. The latest model, Gemini 4 Carbon, was launched two weeks after the release of Gemini 4 Argon for paying subscribers.

Carbon's capabilities and performance

According to a blog post signed by Koray Kavukcuoglu, vice president of Google DeepMind and chief AI architect at Google, Carbon is presented as the most powerful model in the series. Argon remains available for use, although it is not yet accessible in the Gemini application. Carbon is initially reserved for a small number of cybersecurity experts.

Performance comparisons

According to an independent ranking called the Artificial Analysis Intelligence Index, Carbon scores 59 points, surpassing Anthropic's Claude Opus 5.5 (58 points) and OpenAI's GPT-6 Astra (57 points), but outperforming its predecessor, Argon, by only one point. Carbon surpasses GPT-6 Astra and Opus 5.5 on the majority of the published tests, but ranks fourth on WebDev Arena, a web development ranking, despite a gain of 89 Elo points.

Security features and design

Google claims that Carbon has a 0% success rate for prompt injection attacks, compared to 0.7% for Argon on the Gray Swan test. The model is primarily designed for long tasks executed by agents like Spark, which performs tasks in the background in Gemini. Internally, Google's teams have observed Carbon working for several consecutive days to trace the origin of bugs in large code bases and fix them.

Cybersecurity test results

Google emphasizes cybersecurity, with Carbon achieving top scores in three tests: 88.2% on SEC-bench Pro, which involves finding vulnerabilities in real software; 54.8% on ExploitGym, which checks if a vulnerability can be exploited; and 71% on CWE-bench, which measures the ability to write fixes. These figures are from Google.

Access and pricing

The company states that the three cybersecurity tests were conducted "without safeguards," meaning without the filters intended to block malicious requests. However, an AI that can find, exploit, and fix vulnerabilities interests both defenders and hackers. Therefore, Google is opening access to Carbon only for a selection of defenders and plans to expand access to other trusted partners through its Fairwind program.

Pricing and future access

A price is already set for developers, the same as the launch price of Argon: 2 dollars per million tokens sent to the model (about 1.80 euros HT) and 10 dollars per million tokens generated (about 9 euros HT). Once the launch period is over, these prices will double to 4 and 20 dollars (about 3.60 and 18 euros HT). Google has not provided a specific date for opening access to developers, businesses, and the general public, only promising to do so "as soon as possible."

Short release window for Argon

Argon had only two weeks between its announcement on September 30 and its launch for Google AI Pro and Ultra subscribers.