Google launches Gemini 4 Carbon for complex engineering and cybersecurity

A new frontier for engineering

Google has announced Gemini 4 Carbon, its latest frontier model designed to handle complex, long-horizon engineering problems. Unlike models built for quick interactions, Carbon is structured for sustained agentic workflows. This means it can persist through various dead ends to complete multi-step tasks over several days.

The model is specifically tuned to tackle difficult problems like identifying the source of errors and fixing minor glitches within massive software repositories and low-level systems. It builds on the previous Gemini 4 Argon model, which was released just two weeks ago. This rapid succession highlights the pace at which Google is pushing the boundaries of what AI can achieve in software development.

Beyond deep engineering, Carbon shows significant improvements in full-stack web development and multimodal understanding. In the WebDev Arena, the model jumped +89 elo points to reach a ranking of #4 with a score of 1768. These gains are intended to provide higher quality output and artifacts for developers working on complex web projects.

State-of-the-art cyber defense

A primary focus of Gemini 4 Carbon is its performance across the cybersecurity defense lifecycle. It is designed to help defenders find, validate, and remediate vulnerabilities in software. As reported by blog.google, the model achieved a state-of-the-art 88.2% on SEC-bench Pro, where it autonomously uncovered complex flaws in real-world software.

The model also excels at validating risk through exploitation, which helps teams prioritize which bugs to fix first. It set a new record on the ExploitGym benchmark with a score of 54.8%. Additionally, it reached 71.0% on Collinear’s CWE-bench, which measures the model's ability to turn validated findings into actual working code fixes.

Real-world partners are already seeing these results in practice. The Chrome team noted that Carbon was significantly more effective than a rival lab's large frontier model when producing high-quality patches. Furthermore, the cloud security platform Wiz observed Carbon leading on benchmarks for web penetration testing, a key part of their Scan for Good initiative.

Improved alignment and honesty

Google is taking a deliberate approach to safety by training Carbon to respond to high-pressure tasks in an aligned way. The researchers aimed to reduce shortcut attempts during training while maintaining high-quality coding performance. For example, Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, while the actual attempts to cheat fell to 0.4% even as overall accuracy improved.

The model is designed to be more transparent and honest in its communication. Instead of making up information or hedging when it lacks evidence, Carbon now includes methodological caveats in its results or provides a clear list of what is still needed to complete a blocked action. This approach helps cut unsupported claims by approximately 2x compared to the Argon model.

Carbon also handles ambiguity by collaborating more closely with the user. Rather than guessing a path forward when instructions are unclear, it retrieves relevant sources and highlights discrepancies to the user before taking action. This ensures the model stays within project constraints, reducing overreach by 26% and under-service by 29% to provide more reliable execution of multi-step plans.

Rollout and availability

Google is moving forward with a deliberate rollout strategy for Carbon to ensure safety and reliability. The model is currently being shared with a specific group of cyber defenders to gather initial feedback. This phase allows the team to iterate on guardrails before the model reaches a wider audience.

Access will gradually expand to more trusted partners through the Fairwind Program. Once the testing phase is complete, Google plans to make Carbon available to the general public, including businesses, consumers, and developers, via their standard product surfaces. The goal is to provide meaningful leaps in software engineering without compromising on security.

The pricing for Gemini 4 Carbon will match the introductory rates of the Argon model. Users will pay $2 per million input tokens and $10 per million output tokens. Additionally, cached input tokens will be offered at a 95% discount compared to the standard input token price. This pricing structure aims to make the high-performance capabilities of Carbon accessible to a broader range of users.