Google Launches Gemini 4 Carbon, Advanced AI Model Excelling in Cybersecurity and Long-Horizon Coding Tasks

Image: Google · Source

Google unveils Gemini 4 Carbon, a top-tier AI model delivering breakthrough performance in cybersecurity defense and complex coding, with improved safety and alignment features.

Google has announced Gemini 4 Carbon, its newest and most powerful artificial intelligence model to date, designed to advance the frontiers of software engineering and cybersecurity defense. As described by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, Carbon brings state-of-the-art cybersecurity capabilities alongside enhanced ability to manage extended, complex engineering workflows.

Ranked number one on the updated Artificial Analysis Intelligence Index (v5.0) with a score of 59, Gemini 4 Carbon significantly outperforms previous models in key cybersecurity tasks such as vulnerability detection, exploitation validation, and automated patching. It achieved leading scores of 88.2% on SEC-bench Pro for vulnerability finding, 54.8% on ExploitGym for exploitation assessment, and 71.0% on CWE-bench for patching effectiveness. Internally, Carbon demonstrated more than double the success rate in generating secure code compared to Gemini 4 Argon, an earlier model launched two weeks prior.

Developed to address more demanding, multi-step engineering problems, Carbon supports sustained, agentic workflows that can run continuously over days to identify and fix elusive bugs within large, low-level codebases. Improvement in full-stack web development tasks was also noted, with the model scoring 1768 Elo points on WebDev Arena, positioning it fourth among peers.

On the safety and alignment front, Gemini 4 Carbon builds upon Argon's multi-layered defenses, including real-time misuse detection, prompt-injection resilience, and chain-of-thought monitoring, achieving zero successful prompt injection attacks in testing. The model also reduces misleading outputs by approximately half compared to its predecessor and improves collaborative interaction by actively retrieving and presenting relevant information rather than guessing or fabricating answers.

In partnership with select cybersecurity professionals under its Fairwind Program, Google has begun a phased rollout of Carbon to trusted partners, gathering feedback to refine guardrails before wider release. The model is offered at an introductory price matching Argon's, with rates of $2 per million input tokens and $10 per million output tokens, and substantial discounts on cached tokens.

Google emphasizes its deliberate approach to deployment, balancing rapid capability advances with rigorous safety measures. The company plans to expand access to Carbon to developers, businesses, and consumers through various product channels in the near future, anticipating the model will push significant progress in secure, complex software development and cyber defense.