Overview
- Google announced Gemini 4 Carbon on Wednesday, Oct. 28, and said it will provision the model first to a limited cohort of vetted cyber defenders through its Fairwind partner program.
- Google published benchmark scores showing Carbon at 59 on the Artificial Analysis Intelligence Index and leading cybersecurity results: 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench.
- On internal tests Google reported safety and reliability gains, including more than a twofold uplift in secure-code generation over Argon, a near‑zero indirect prompt‑injection success rate, and roughly half the unsupported claims versus Argon.
- The company said it is engaging the U.S. voluntary pre-release testing process, will phase wider access after gathering feedback, and set introductory token pricing at $2 per 1M input tokens and $10 per 1M output tokens with deep discounts for cached inputs.
- Independent third‑party verification of Google’s performance and safety claims remains limited, a gap that will shape how quickly broader use proceeds and raises questions about dual‑use risks and who can afford sustained agentic cyber workflows.