Google introduces Gemini 4 Carbon, with cybersecurity testing planned first

Google has introduced Gemini 4 Carbon with claims of stronger coding capabilities and cybersecurity performance. Selected companies working in cybersecurity defense will receive access first for testing before release, with broader availability planned after additional partners are admitted.

The results Google published put Carbon at 59 on AA Index v5.0, a benchmark assessing general AI capabilities. That score placed it slightly ahead of Claude Opus 5.5 and GPT-6 Astra. Google reported wins in nearly all benchmark categories, spanning math, science, knowledge work, computer use and agentic coding.

For cybersecurity tasks, Google said the model can locate vulnerabilities, determine their priority and fix the underlying problems. The company also said Carbon has twice the likelihood of generating secure code as Gemini 4 Argon. Other claimed improvements include greater resistance to prompt injection, better collaboration and transparency, and less reliance on shortcuts or loopholes.

Google plans to extend access to more trusted partners through its Fairwind Program before a wide release. It used that approach for Gemini 4 Argon, which users can now access through the Gemini app and AI search mode.