Google DeepMind has unveiled Gemini 4 Carbon, its most powerful AI model yet, designed for cybersecurity, programming, and complex multi-step tasks. The model can find vulnerabilities in real software, test for exploitability, and create patches, according to the company.
Google DeepMind has unveiled Gemini 4 Carbon, its most powerful AI model yet, designed for cybersecurity, programming, and complex multi-step tasks. The model can find vulnerabilities in real software, test for exploitability, and create patches, according to the company.
The Gemini 4 Carbon will initially be made available to a limited number of cybersecurity professionals. Google will then gradually expand access through the Fairwind program to verified partners. The company also said it is participating in a voluntary pre-release testing process coordinated by the U.S. government.
The launch was announced on October 28 by Google DeepMind Senior Vice President and Google Chief AI Architect Koray Kavukcuoglu.
According to Google, Gemini 4 Carbon ranked first in the updated version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The company also claims significant improvements in programming, AI agent work, and scientific tasks.
In WebDev Arena, a ranking of web application development models, Carbon ranked fourth with a score of 1768 points. This is an increase of 89 Elo points compared to the previous version of the model.
Google is positioning Carbon as a model for long-term engineering tasks that require multiple steps to be performed sequentially. According to internal company observations, the model can work on complex tasks for several days, including finding the causes of subtle errors and fixing them in large code bases and low-level systems.
Carbon builds on the capabilities of Gemini 4 Argon, which Google announced two weeks ago. The company says it has taken user feedback into account and improved its support for programming, agent-based scenarios, scientific tasks, web development, and multimodal data.
One of the main features of the Gemini 4 Carbon is its performance in cybersecurity tests. Google says that the model demonstrates high performance in various stages of software protection: from vulnerability detection to risk verification and preparation of patches.
According to the company, Carbon showed the following results:
88.2% in SEC-bench Pro — a test for detecting vulnerabilities in real software;
54.8% in ExploitGym, a benchmark for testing the possibility of exploiting vulnerabilities;
71% in CWE-bench, a test that assesses the ability to find root causes of problems and prepare working fixes.
Google also reports that in internal tests, Carbon more than doubled the success rate of generating secure code compared to Gemini 4 Argon.
The Chrome browser team, according to the company, received more than ten times better results from Carbon in terms of the quality of prepared patches compared to another leading model from an unnamed AI lab. Cloud cybersecurity platform Wiz also reported Carbon’s leadership in web penetration tests.
At the same time, Google clarifies that external cybersecurity benchmarks were conducted without the model’s protective constraints. This is an important nuance for evaluating its capabilities in scenarios where AI can be used both to protect systems and to find attack vectors.
Special attention was paid to safety and reliability during the development of the Gemini 4 Carbon. Google claims that the model is less likely to make up facts, try to circumvent restrictions, or perform actions outside the scope of the task at hand.
According to the company’s internal tests, the number of unsupported claims in Carbon is about half that of Argon. If evidence or necessary data is lacking to complete a task, the model should point out the limitations rather than invent the answer.
In Terminal-Bench 4.0 testing, the rate of successful fraud attempts—including attempts to circumvent the evaluation rules—was 0%, and the rate of attempts themselves dropped to 0,4%. According to observations by internal Google teams, the number of warnings about serious problems with model consistency behavior decreased by almost 80%.
Carbon also showed a 0% success rate in the Gray Swan Indirect Prompt Injection benchmark, which tests whether a model can withstand hidden instructions that attackers try to pass through external data sources.
Google notes that Carbon uses a multi-layered protection system inherited from Argon. It includes model training, real-time abuse protection, prompt injection resistance, and monitoring of internal reasoning processes. The company emphasizes that it plans to develop the model’s capabilities in parallel with strengthening the safeguards.
Google is not opening the Gemini 4 Carbon to everyone at once. It will initially be available to select cybersecurity professionals, with access gradually expanding through the Fairwind program to verified partners.
The company plans to continue gathering feedback and improving security mechanisms before launching it for developers, businesses, and consumers. Google has not yet announced a specific date for the public release.
The starting price of the Gemini 4 Carbon will be the same as the Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens are discounted by 95% from the standard price.
