Google launches Gemini 4 Carbon, topping AI rankings with top security skills

Google has launched its new frontier model Gemini 4 Carbon, built for top-tier cybersecurity and long-running tasks. It ranked first on the latest Artificial Analysis index with a score of 59. Launch pricing matches Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off.

The model is designed for agent workflows that run for days. It can find root causes in large codebases and write patches on its own. It scored 88.2% on the SEC-bench Pro vulnerability discovery test and 71.0% on CWE-bench root cause fixing, and its safe code success rate is over 2 times that of Argon.

On safety, its cheating success rate and indirect prompt injection attack success rate both dropped to 0.0%, with made-up claims cut by about half. Google will first open access to a small group of cybersecurity defenders, then expand through the Fairwind program before releasing it to developers, businesses, and consumers.