Gemini 4 Carbon has been announced as Google's latest and most robust model, featuring cutting-edge cybersecurity and extended horizon capabilities.
Carbon has achieved the top spot on the updated v5.0 Artificial Analysis Intelligence Index with a score of 59. It demonstrates notable improvements in coding and agentic tasks, excelling in key areas essential for cybersecurity. The model leads in SEC-bench Pro with an 88.2% success rate for identifying vulnerabilities, scores 54.8% in ExploitGym for exploitation tasks, and achieves 71.0% in CWE-bench for patching efforts. Internally, Carbon shows a strong capability in generating secure code and reducing misalignment and fabrication in live coding environments.
This new model builds on the Gemini 4 Argon, which was released two weeks prior. Carbon highlights the rapid progress being made in advancing technology while enhancing safeguards. Google is carefully rolling out the model, participating in the U.S. government's voluntary pre-release testing process. Initially, Carbon will be accessible to a select group of cyber defenders, with plans to gradually extend access to additional trusted partners through the Fairwind Program. Feedback will be collected to refine guardrails before the model is widely available to developers, businesses, and consumers.
The introductory pricing for Gemini 4 Carbon matches Argon's, at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens.
Carbon is designed to tackle complex, long-duration engineering challenges, sustaining agentic workflows over extended periods. Our teams report that Carbon can operate for days, navigating dead ends to resolve intricate, multi-step engineering tasks such as diagnosing and fixing subtle bugs in large codebases and low-level systems.
Responding to user feedback, improvements have been made in model performance across diverse areas, including coding, agentic tasks, and scientific endeavors, as well as full-stack web development and multimodal understanding. With these enhancements, Carbon has increased by 89 elo points, now ranking fourth on WebDev Arena with a score of 1768.
Gemini 4 Carbon represents a major advancement in cybersecurity capabilities, delivering cutting-edge performance throughout the cybersecurity defense lifecycle, from vulnerability detection to validation and remediation. Internal evaluations show Carbon achieves more than double the success rate of Argon in generating secure code. The Chrome team also noted that Carbon outperformed another lab's large frontier model by more than 10 times in creating quality patches. Additionally, Wiz, Google's Cloud security platform, recognizes Carbon's leadership in web penetration testing, a key component of Wiz's Scan for Good initiative.
Carbon builds on the multi-layered safety and security architecture of Argon, incorporating model training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. With significant advancements in model alignment, Carbon is trained to handle complex or high-pressure tasks effectively.
As we push the boundaries of technology, moving swiftly yet carefully is crucial for deployment. With Gemini 4 Carbon, we aim to make substantial advancements in software engineering and cybersecurity without sacrificing safety or reliability. We are excited to observe how the initial group of cyber defenders utilizes this model and look forward to making Carbon available to developers, businesses, and consumers in the near future.