Google Launches Gemini 4 Carbon With Big Cybersecurity Gains

Google has announced Gemini 4 Carbon, a new AI model with major improvements in coding, autonomous tasks and cybersecurity. In an announcement on Wednesday, the company said the model leads the Artificial Analysis Intelligence Index (version 5.0), ahead of Claude Opus 5.5 and GPT-6 Astra.

Gemini 4 Carbon succeeds Argon, which Google released earlier this month. It is part of the company's push toward models that can handle complex engineering work and strengthen cyber defenses.

Security benchmarks

According to Google, the model can find vulnerabilities in software and write more secure code. It topped SEC-bench Pro, which measures how well AI models find real bugs and flaws in software, and ExploitGym, which tests whether they can turn discovered vulnerabilities into working exploits. Google also cited results showing it beat GPT-6 Astra on SEC-bench Pro.

Long tasks and reliability

The model is built for long, multi-step tasks and is meant to keep working when it hits obstacles. Google said its internal teams tested it on engineering jobs lasting days, and that it got past dead ends to keep working on hard problems.

The company also says the model is better at avoiding inappropriate shortcuts, resisting prompt injection attacks that try to manipulate AI behavior, and collaborating with users. Google adds that it cuts unsupported claims by half compared with Argon, which should improve the accuracy and transparency of its answers.

Availability and pricing

For now, Gemini 4 Carbon is available only to a limited group of cyber defense companies. Google plans to widen access later to additional partners through its Fairwind program.

Pricing starts at $2 per million input tokens and $10 per million output tokens, the same introductory rates Google used when it launched Argon.