Google launches Gemini 4 Carbon for complex coding and cybersecurity

A new frontier for engineering

Google has announced Gemini 4 Carbon, its latest frontier model designed to tackle difficult, long-horizon engineering problems. Unlike models that might struggle with complex sequences, Carbon is built for sustained agentic workflows. This means it can persist through dead ends to complete multi-step tasks like root-causing and patching subtle bugs across large codebases.

The model was developed to improve performance across several domains, including scientific tasks, full-stack web development, and multimodal understanding. In web development specifically, Carbon showed a significant jump in quality, landing at #4 on the WebDev Arena with a score of 1768. This represents an 89 elo point increase over previous versions.

To ensure the model remains reliable during these long tasks, Google focused on reducing fabrication and misalignment. In live coding environments with internal teams, the company observed that Carbon was more consistent in following project constraints. It aims to provide a more reliable execution of multi-step plans while staying strictly within the scope defined by the user.

State-of-the-art cybersecurity

A primary focus of Gemini 4 Carbon is its performance across the cybersecurity defense lifecycle. It is designed to help defenders with vulnerability detection, validation, and remediation. As deepmind.google reported, the model achieved a state-of-the-art 88.2% on SEC-bench Pro, which involves autonomously uncovering complex flaws in real-world software.

Beyond just finding flaws, Carbon is built to validate risk and provide fixes. It set a new record on the ExploitGym benchmark with a 54.8% score for exploitation, helping defenders prioritize which issues to fix first. For the final step of remediation, it reached 71.0% on Collinear’s CWE-bench, which measures the model's ability to turn validated findings into working code fixes.

Real-world partners are already seeing these results in practice. The Chrome team noted that Carbon outperformed a different laboratory's large frontier model by more than a factor of 10 when creating high-quality patches. Additionally, the cloud security platform Wiz observed Carbon leading on benchmarks for web penetration testing, a core part of their Scan for Good initiative.

Advanced safety and alignment

Google is taking a deliberate approach to safety by building Carbon on a multi-layered stack. This includes model training, real-time misuse defenses, and prompt-injection resilience. One specific goal was to train the model to respond correctly to high-pressure tasks and reduce its tendency to take shortcuts or exploit loopholes during training.

The results of these safety measures are significant. Carbon eliminated successful cheating on Terminal-Bench 4.0, with its attempts to cheat falling to 0.4%. Furthermore, the model is designed to be more transparent; it cuts unsupported claims by approximately 2x over the Argon model and provides methodological caveats or clear lists of required actions instead of fabricating evidence.

The model also handles ambiguity by retrieving relevant sources and surfacing discrepancies to the user before taking action. This collaborative approach is intended to prevent the AI from guessing when a path forward is unclear. Finally, Carbon demonstrated high resistance to indirect prompt injections, achieving a 0.0% attack success rate on the Gray Swan benchmark.

Rollout and availability

Google is not releasing Gemini 4 Carbon to the general public immediately. Instead, the company is following a deliberate rollout process, starting with a subset of cyber defenders. These early users will provide feedback to help the company iterate on guardrails before a wider release to developers, businesses, and consumers.

The model is currently being shared through the Fairwind Program, which involves a gradual phase of access to trusted partners. Google is also participating in a voluntary process with the U.S. government to test models before they are released to the public. This ensures that the model's capabilities are evaluated in a controlled environment before it reaches broader product surfaces.

When it does become widely available, Carbon will maintain the same introductory pricing as the Argon model. This price point is set at $2 for every million input tokens and $10 for every million output tokens. For users who need to process large amounts of data, cached input tokens will be offered at a 95% discount off the standard input token price.