Google has announced its new flagship AI model, Gemini 4 Carbon. Positioned as the company's most powerful model to date, Carbon focuses specifically on multi-step, long-duration software tasks and cybersecurity operations.
Google CEO Sundar Pichai made the announcement on his social media account. Recalling his previous statement that they would release models at regular intervals, Pichai noted that Carbon follows Gemini 4 Argon, which was opened to general use just two weeks ago. This schedule highlights the pace of the model development race among major AI companies.
The model reached the top spot with 59 points in the updated Intelligence Index v5 ranking by the independent evaluation organization Artificial Analysis. In the same ranking, Anthropic's Claude Opus 5.5 model scored 58 points, while OpenAI's GPT-6 Astra model scored 57. However, Carbon lagged behind Claude Opus 5.5 in Artificial Analysis's coding agent index.
DESIGNED FOR LONG TASKS AND CODEBASES
According to Google DeepMind, Gemini 4 Carbon was developed for AI agents capable of executing complex workflows that can last for hours or even days, rather than just generating one-off responses. Internal company tests reportedly showed that the model could find and fix the root causes of hidden bugs in large codebases.
Google states that Carbon has also made progress in coding, scientific tasks, web development, and understanding visual content. It was reported that the model rose to 4th place with an 89-point increase in the WebDev Arena ranking, which measures web development capabilities.
On the cybersecurity front, the company claims that Carbon has delivered strong results in a defense chain that extends from detecting a vulnerability to verifying its exploitability and generating a permanent patch. It was announced that the model scored 88.2 percent in the SEC-bench Pro test, 54.8 percent in ExploitGym, and 71.0 percent in CWE-bench.
According to Google, Carbon showed more than twice the success of Argon in secure code generation. In an evaluation by the Chrome team, it was stated that the model produced 10 times better results in high-quality patch generation compared to another unnamed major model. However, it is important to note that external tests were conducted with the model's security filters disabled.
Gemini 4 Carbon is initially being opened to selected cybersecurity experts under Google's Fairwind Program. It was reported that access will later be expanded to trusted partners, and it will be offered to developers, companies, and end-users once security measures are matured through feedback. Google did not share a clear date for this general access.
The developer pricing for the model will be at the same level as Gemini 4 Argon during the introductory period: 2 dollars per 1 million input tokens and 10 dollars per 1 million output tokens. A 95 percent discount will be applied to previously processed cached inputs. After the introductory period, prices are expected to rise to 4 dollars and 20 dollars, respectively.
According to Artificial Analysis, Carbon offers a 1 million token context window. This capacity allows the model to work on very long documents or large software projects at once. Google has not announced a separate access schedule for countries including Turkey; users in Turkey can currently access the generally available Gemini 4 Argon.
Initial evaluations of the model point not only to strong results but also to some risks. Artificial Analysis stated that Carbon's hallucination rate increased from 4 percent in Argon to 11 percent. Andon Labs shared that in their own tests, Carbon exhibited some problematic behaviors more mildly than Argon, but lagged behind GPT-6 Astra in this area.
While Carbon's most prominent advantages stand out in cybersecurity, long context, and cost, it is observed that competing models retain some superiority in classic software development and computer usage tests. The model's true performance will be clearer once the limited access period ends and it is tested by broader developer communities.
News Source: 12punto
