Google Unveils Gemini 4 Carbon, Its Most Advanced AI Model for Cybersecurity

Google has unveiled Gemini 4 Carbon, a new frontier artificial intelligence model developed by Google DeepMind to tackle complex software engineering tasks and strengthen cybersecurity. Announced on October 28, 2026, the model arrives shortly after the introduction of Gemini 4 Argon, signalling Google’s determination to compete more aggressively with leading AI developers such as OpenAI and Anthropic. According to Google, Carbon delivers significant improvements in coding, reasoning and agentic capabilities, allowing it to work through complicated tasks that may require numerous steps and sustained effort. The company says the model can help engineers identify subtle software defects, understand problems across large codebases and develop solutions that address the underlying causes of technical failures. These capabilities could make advanced AI increasingly useful in professional software development, where reliability and security are just as important as generating code quickly.

One of Gemini 4 Carbon’s most significant features is its reported performance in cybersecurity. Google says the model achieved an 88.2% score on SEC-bench Pro, a benchmark designed to evaluate the discovery of software vulnerabilities. It also recorded 54.8% on ExploitGym, which evaluates the ability to determine whether vulnerabilities can be exploited, and 71.0% on CWE-bench, which assesses vulnerability remediation. These results address several stages of the security process: finding weaknesses, assessing their practical significance and developing patches to fix them. Google further reports that Carbon generated secure code at more than twice the success rate of Gemini 4 Argon on its internal evaluations. If these improvements translate into real-world performance, the technology could help security teams investigate software flaws faster, prioritise urgent fixes and reduce the time systems remain exposed to known weaknesses. However, these figures are largely based on benchmark testing and company-reported evaluations, so independent assessments will be important in determining how well the model performs across different environments.

Beyond cybersecurity, Carbon is designed to handle long-running engineering projects and agentic workflows. Rather than simply responding to a single instruction, an AI agent can plan a sequence of actions, use available tools and continue working towards a defined objective. Google says its internal teams have observed Carbon persisting through difficult, multi-step engineering problems, including diagnosing and correcting subtle bugs across large software systems. The company also reports improvements in full-stack web development, scientific work and multimodal understanding. On WebDev Arena, Google says Carbon gained 89 Elo points, reaching fourth place with a score of 1,769. Such capabilities could be valuable to software developers, technology companies and research teams seeking to automate repetitive work while concentrating human expertise on more demanding decisions. Nevertheless, complex AI-driven workflows still require oversight, testing and clear limits on what systems are permitted to change, particularly when they interact with production software or sensitive data.

Google is also placing considerable emphasis on safety as AI models become more capable of identifying and exploiting software vulnerabilities. The company says Carbon includes safeguards involving model training, real-time misuse prevention, resistance to prompt injection and monitoring of model reasoning. Prompt injection is a type of attack in which malicious instructions are introduced through content an AI system processes, potentially causing it to ignore its intended task or violate restrictions. Google reports that Carbon resisted all evaluated attacks in the Gray Swan indirect prompt-injection benchmark, recording a 0% attack-success rate in that test. It also says the model reduced unsupported claims and showed improvements in following instructions without taking unauthorised actions. Despite these results, no single benchmark can establish that an AI system is completely secure. Google has therefore adopted a staged rollout, beginning with a selected group of cybersecurity defenders and expanding access through its Fairwind programme as testing continues. The company is also participating in a voluntary US government process for evaluating advanced models before wider release.

The model’s pricing and restricted availability are also important parts of Google’s announcement. During its introductory pricing period, Gemini 4 Carbon is set to cost $2 per million input tokens and $10 per million output tokens, matching the introductory rates announced for Gemini 4 Argon. Google says cached input tokens will receive a 95% discount against the input-token price, while standard rates are expected to rise to $4 per million input tokens and $20 per million output tokens after the introductory period. However, Carbon is not immediately available to everyone: Google says it will initially provide access to selected cybersecurity professionals before gradually extending access to trusted partners and, eventually, developers, businesses and consumers. The announcement places Google in a competitive position as AI companies increasingly focus on models capable of carrying out sophisticated work rather than simply generating text. For businesses, the central question will be whether Carbon’s reported gains translate into measurable improvements in security, productivity and cost efficiency. Its wider release and independent real-world evaluations should help establish whether the model represents a meaningful advance in practical AI engineering.