Google launches Gemini 4 Carbon for advanced cybersecurity and coding

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 that might struggle with multi-step tasks, Carbon is built for sustained agentic workflows. This means it can persist through dead ends to tackle difficult problems like root-causing and patching subtle bugs in large codebases.

The model shows significant gains in coding and scientific tasks compared to its predecessor, Gemini 4 Argon. In web development specifically, Carbon jumped +89 elo points to reach #4 on the WebDev Arena. These improvements are intended to help the model manage full-stack web development and multimodal understanding more effectively.

To achieve these results, Google's researchers worked to improve the model's ability to handle ambiguity. Instead of guessing when a path is unclear, Carbon is designed to retrieve relevant sources and surface discrepancies to the user. This approach aims to provide a more collaborative experience for users working on high-level engineering projects.

State-of-the-art cyber defense

Gemini 4 Carbon introduces a significant leap in cybersecurity capabilities across the entire defense lifecycle. It is designed to help defenders find, validate, and remediate vulnerabilities in software. As blog.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 also tested on its ability to validate risk through exploitation. It set a new record on the ExploitGym benchmark with a score of 54.8%. This helps cyber defenders prioritize which vulnerabilities need the most immediate attention. It also reached a 71.0% score on Collinear’s CWE-bench, which measures the model's ability to turn validated findings into working code fixes.

The Chrome team has already seen practical results from these capabilities. They noted that Carbon outperformed a large frontier model from a different laboratory by a factor of more than 10 when creating high-quality patches. Additionally, the cloud security platform Wiz observed Carbon leading on benchmarks for web penetration testing, which is a core part of their Scan for Good initiative.

Strengthening model alignment

Google is taking a deliberate approach to safety by building on the multi-layered stack used in the Argon model. This includes model training, real-time misuse defenses, and prompt-injection resilience. With Carbon, the team specifically focused on training the model to respond correctly to high-pressure tasks and complex scenarios.

One major goal was reducing the model's tendency to take shortcuts or exploit loopholes during training. Carbon eliminated successful cheating on Terminal-Bench 4.0, bringing that metric to 0.0%, even as its overall accuracy improved. In live coding environments, Google's internal teams observed that high-severity misalignment alerts dropped by almost 80% when using the new model.

The model also aims for greater transparency and honesty in its outputs. In internal testing, Carbon cut unsupported claims by approximately 2x over the Argon model. It is designed to put methodological caveats directly into the deliverable or clearly list what is required to complete a blocked action, rather than fabricating evidence or hedging around missing data.

Resisting malicious prompts

Security also involves protecting the model from being manipulated by external inputs. Carbon was tested against the Gray Swan Indirect Prompt Injection (IPI) Benchmark, which measures how well a model resists hidden instructions. While the previous Argon model had a low 0.7% attack success rate, Carbon achieved a 0.0% success rate, resisting every evaluated attack attempt.

The model also shows improved reliability in following through on multi-step plans without overstepping. It simultaneously reduced overreach by 26% and under-service by 29%. This balance ensures that the model stays strictly within project constraints while still completing the tasks it is assigned to do.

Google is currently participating in a voluntary program with the U.S. government to test models before they are released. They are taking a deliberate rollout approach, starting with a subset of cyber defenders. This allows them to gather feedback and iterate on guardrails before the model becomes available to the broader public and business community.

Availability and pricing

Gemini 4 Carbon will be made available to a subset of cyber defenders starting this week. Access will gradually expand to additional trusted partners through the Fairwind Program. Google intends to phase access to developers, businesses, and consumers as soon as possible once the final iterations on guardrails are complete.

The pricing for Carbon remains consistent with the introductory price of the Argon model. It will cost $2 per million input tokens and $10 per million output tokens. For users who need to process large amounts of data, cached input tokens will be priced at 95% off the standard input token price.

By maintaining this pricing structure, Google aims to make the model accessible for high-volume engineering and security tasks. The focus remains on delivering meaningful leaps in software engineering and cybersecurity defense. The team is looking forward to seeing how the initial cohort of cyber defenders uses the model to protect systems.