Google DeepMind has introduced Gemini 4 Carbon, its latest frontier AI model, with a particular focus on cybersecurity, advanced software engineering, and complex multi-step tasks. The new model builds on Gemini 4 Argon, which launched two weeks earlier, and aims to help security researchers identify vulnerabilities, validate potential exploits, and develop fixes for software weaknesses.

Announced on October 28, 2026, Gemini 4 Carbon is Google’s strongest model to date, according to the company. It also brings improvements in coding, reasoning, and agentic workflows, allowing it to work through challenging engineering problems that can take days to complete.

However, Google is taking a cautious approach to its release, initially making the model available to a limited group of cybersecurity professionals before expanding access to other users.

Gemini 4 Carbon puts cybersecurity at the forefront

One of the biggest improvements in Gemini 4 Carbon is its ability to assist throughout the cybersecurity defense lifecycle, covering vulnerability discovery, exploit validation, and remediation.

Google reports that Carbon achieved an 88.2% score on SEC-bench Pro, a benchmark focused on identifying real-world software vulnerabilities. It also scored 54.8% on ExploitGym, which evaluates the ability to validate vulnerabilities through exploitation, and 71.0% on CWE-bench, which measures how effectively a model can address software weaknesses with code fixes.

These results position Carbon among the leading models in the evaluated cybersecurity benchmarks, according to Google.

The company also highlighted improvements in its internal testing. Carbon achieved more than twice the success rate of Gemini 4 Argon when generating secure code. Google’s Chrome team additionally reported that Carbon performed more than 10 times better than another unnamed frontier AI model when generating high-quality software patches.

Google’s cloud security platform, Wiz, also observed strong performance from Carbon in web penetration testing benchmarks, supporting its Scan for Good initiative.

These capabilities could make AI a more useful tool for security teams. Rather than simply identifying a potential vulnerability, a model can help investigate whether it is exploitable and work toward a practical fix.

Better coding and long-running agentic workflows

Cybersecurity isn’t the only area where Google is positioning Carbon as a major upgrade. The model is also designed to handle demanding software engineering tasks that require extended reasoning and sustained execution.

According to Google, its internal teams have tested Carbon on complex engineering problems that involve investigating subtle bugs across large codebases and low-level systems. The model can continue working for days, navigating dead ends and attempting different approaches before reaching a solution.

Google has also reported improvements across coding, scientific tasks, agentic workflows, full-stack web development, and multimodal understanding.

On WebDev Arena, Carbon gained 89 Elo points to reach fourth place with a score of 1,768. This reflects Google’s efforts to improve the quality of generated websites and other development-related outputs.

The broader objective is to make Gemini more capable of handling complete engineering workflows instead of simply responding to individual coding prompts.

Google strengthens Gemini’s safety and security protections

Giving an AI model advanced cybersecurity capabilities also creates additional risks. The same capabilities that help defenders identify weaknesses could potentially be misused to discover and exploit vulnerabilities.

Google says Gemini 4 Carbon builds on the multi-layered safety framework introduced with Argon. This includes model training, real-time misuse defenses, protection against prompt injection, and chain-of-thought monitoring.

The company has also worked on making Carbon more reliable when handling complicated instructions and high-pressure tasks.

In internal testing, Google reported several improvements:

• Fewer attempts to exploit shortcuts: Carbon reduced successful cheating on Terminal-Bench 4.0 to 0%, with attempts to cheat falling to 0.4%, while improving overall accuracy.
• Better alignment: High-severity misalignment alerts dropped by almost 80% in Google’s internal live coding environments.
• More reliable answers: Carbon reduced unsupported claims by approximately two times compared with Argon.
• Improved task execution: The model reduced overreach by 26% and under-service by 29%, helping it follow instructions without exceeding the requested scope.
• Stronger prompt-injection resistance: Carbon recorded a 0% attack success rate on the Gray Swan Indirect Prompt Injection benchmark, resisting every evaluated attack attempt.

These figures come from Google’s internal evaluations and should be understood in that context. They demonstrate the improvements the company reports, rather than guaranteeing that the model will be free from errors or security risks in real-world use.

Gemini 4 Carbon availability and pricing

Google isn’t opening Gemini 4 Carbon to everyone immediately. The company is taking a gradual deployment approach, beginning with a selected group of cybersecurity defenders.

The initial rollout will take place through Google’s Fairwind Program, which provides access to trusted cybersecurity partners. Google also says it is participating in the US government’s voluntary process for pre-release model testing.

The company plans to expand access to developers, businesses, and consumers through its products once it has gathered sufficient feedback and refined the model’s safeguards.

For developers using the model through its API, Google has announced the same introductory pricing as Gemini 4 Argon:

• Input tokens: $2 per million tokens.
• Output tokens: $10 per million tokens.
• Cached input tokens: 95% less than the standard input token price.

After the introductory pricing period expires, the standard rates will increase to $4 per million input tokens and $20 per million output tokens.

For now, Gemini 4 Carbon represents another step in Google’s efforts to develop AI systems that can do more than generate code. Its combination of cybersecurity research, vulnerability remediation, and long-running engineering workflows points toward a future where AI assistants can take a more active role in maintaining and securing complex software systems.