Google unveils Gemini 4 Carbon AI model with record-breaking cybersecurity capabilities

It notched record-breaking benchmark scores.88.2% on SEC-bench Pro, autonomously finding complex flaws in real-world software.71.0% on CWE-bench, translating validated findings directly into secure code patches.

Google's Chrome team noted that Carbon generated high-quality security patches more than 10 times better than competing frontier models.Designed to manage complex, long-running engineering chores over multi-day spans, Carbon significantly minimizes hallucinations, fabrications, and task overreach/under-service.Google tightened safety mechanisms alongside the model's increased capability.

Internal testing demonstrated that Carbon brought successful cheating attempts on Terminal-Bench down to effectively 0%, while maintaining stricter adherence to scope and project constraints.Unlike broader consumer rollouts, initial access to Carbon is strictly gated.

It is rolling out to a curated subset of cyber defenders and trusted partners through Google’s Fairwind Program, with wider availability planned later.When accessible via APIs, Google has slated introductory pricing of $2 per million input tokens and $10 per million output tokens (with cached input tokens receiving a 95% discount), matching the pricing tier of its Argon counterpart.

Google unveils Gemini 4 Carbon AI model with record-breaking cybersecurity capabilities October 29, 2026.