Google Unveils Gemini 4 Carbon Frontier Model
New AI model prioritizes cybersecurity defense and long-horizon engineering capabilities.
1Introducing Gemini 4 Carbon
Google DeepMind announced Gemini 4 Carbon, a frontier model optimized for cybersecurity and long-horizon engineering. Released on October 28, 2026, it follows the Gemini 4 Argon launch by only two weeks. The model currently ranks first on the Artificial Analysis Intelligence Index v5.0, demonstrating significant advancements in coding and agentic task execution.
2Cybersecurity Performance Metrics
Carbon leads in SEC-bench Pro with 88.2% for vulnerability detection, 54.8% in ExploitGym, and 71.0% in CWE-bench for patching. Internal benchmarks show a 2x success rate uplift in secure code generation compared to Argon. The Chrome team reported the model performed 10x better than competing frontier models when generating high-quality security patches.
3Long-Horizon Engineering Capabilities
Designed for sustained agentic workflows, Carbon handles complex, multi-step engineering tasks that require days of persistence. It excels at root-causing and patching subtle bugs within large codebases and low-level systems. This capability allows the model to navigate dead ends effectively, providing a robust solution for full-stack web development and complex software engineering challenges.
4Web Development Benchmarking
The model demonstrates substantial improvements in web development tasks, achieving an 89-point Elo increase on the WebDev Arena. This performance jump places Carbon at the fourth position with a score of 1768. These gains reflect a focused effort to enhance output quality and artifacts, supporting developers in building more reliable and secure web applications.
5Safety and Alignment Strategy
Carbon utilizes the same multi-layered safety stack as Argon, incorporating real-time misuse defenses and prompt-injection resilience. The training process emphasized alignment, specifically teaching the model to maintain correct behavior during high-pressure or complex tasks. This approach ensures that the leap in capability does not compromise the overall reliability of the system.
6Phased Deployment Approach
Google is taking a deliberate rollout approach, engaging in U.S. government voluntary pre-release testing. Initial access is limited to a subset of cyber defenders, with plans to expand via the Fairwind Program. Feedback gathered during this phase will inform final guardrails before the model is released to broader developer, business, and consumer audiences.
7Pricing and Economic Structure
Gemini 4 Carbon launches with an introductory price of $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount. Once the introductory period expires, standard pricing will double to $4 per million input tokens and $20 per million output tokens for all users.
8Strategic Industry Impact
By integrating advanced cybersecurity defense with agentic reasoning, Carbon represents a shift toward AI that actively participates in software maintenance. Its performance in penetration testing, validated by partners like Wiz, highlights its utility for enterprise security. The rapid release cycle underscores Google's commitment to pushing frontier capabilities while maintaining strict safety protocols.
Original article: DeepMind — Google Unveils Gemini 4 Carbon Frontier Model