Google announced Gemini 4 Carbon on October 28, 2026, describing it as the company’s most powerful frontier model to date. However, Carbon is not coming with a broad release. Initial access is going to a select group of cyber defenders, not developers or end users.
Phased rollout and the Fairwind Program
Google DeepMind Senior Vice President and Google’s Chief AI Architect Koray Kavukcuoglu described Carbon as the company’s most powerful model. The model is credited with state-of-the-art cybersecurity and long-horizon task capabilities. Carbon replaces Gemini 4 Argon, announced two weeks ago.
The company says it has included the model in the U.S. government’s voluntary pre-release testing process. Starting in the announcement week, Carbon is being opened to a limited number of cyber defenders; access will then be gradually expanded to trusted partners through the Fairwind Program. Google DeepMind reports that it works with more than 650 partners worldwide in the program.
The program’s terms are strict. Partners may only perform defensive dual-use work such as authorized threat simulation, reverse engineering, and malware analysis; producing malware is prohibited. Organizations must implement user-based authentication and phishing-resistant multi-factor authentication, and limit access to internal cybersecurity, incident response, and penetration testing teams. Access cannot be shared or sold; applicants’ security history and ethical business record are also audited.
Coding and test results on long tasks
According to Google, Carbon is designed for long-running engineering problems and targets agent workflows that last for days. It is reported that internal teams have run the model for days on multi-step tasks such as finding and patching the root cause of subtle bugs in large codebases.
It was announced that it ranks first on version 5.0 of the Artificial Analysis Intelligence Index with 59 points. In WebDev Arena, it rose to fourth place with 1768 points after gaining 89 Elo points. Google says it focused its improvements on coding, agent tasks, scientific work, full-stack web development, and multimodal understanding.
Measured results in cyber defense
The external evaluations shared by the company were conducted with guardrails disabled. Scores of 88.2 were reported on SEC-bench Pro, which measures autonomously finding complex vulnerabilities in real software; 54.8 on ExploitGym, which verifies the exploitability of discovered vulnerabilities; and 71.0 on CWE-bench, which tests patches from the root cause. The ExploitGym result is presented as a new record.
In Google’s internal measurements, the model delivered more than a twofold increase in success rate for secure code generation compared with Argon. It is reported that the Chrome team found Carbon to produce more than 10 times better results than another lab’s large frontier model in high-quality patch generation. Cloud security platform Wiz also sees Carbon as the leader in web penetration testing tests.
Security, pricing, and access plan
Carbon builds on Argon’s security stack: model training, real-time abuse defenses, prompt injection resistance, and chain-of-thought monitoring.
Google reports that on Terminal-Bench 4.0, the successful cheating rate dropped to zero, cheating attempts remained at 0.4%, and accuracy improved. In live coding environments, high-severity alignment warnings fell by about 80%. Unsupported claims fell by roughly half compared with Argon. In multi-step plan execution, overreach decreased by 26% and underservice by 29%. In Gray Swan’s indirect prompt injection test, the attack success rate was 0%; Argon had led this test with 0.7%.
Pricing starts at the same level as Argon: $2 per million input tokens and $10 per million output tokens; cached input tokens receive a 95% discount. When the introductory period ends, prices will rise to $4 per million input tokens and $20 per million output tokens.
Google says it plans to open Carbon to developers, businesses, and consumers as soon as possible.
Source: Unite.AI
Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.