Google Announces Gemini 4 Carbon, Its Latest Frontier AI Model That You Can't Use
Gemini 4 Carbon is essentially a slightly better version of Argon.
TL;DR
- Google has released Gemini 4 Carbon, an updated AI model following Gemini 4 Argon.
- Carbon shows significant improvements in coding, agentic tasks, and scientific analysis, especially in cybersecurity applications like vulnerability detection and remediation.
- Google claims Carbon is twice as likely to generate vulnerability-free code compared to Argon and 10 times better than non-Google models in generating high-quality patches.
- The new model ranks fourth on the WebDev Arena leaderboard with an 89-point score increase.
- Carbon reportedly outperformed competitors like GPT-6 Astra and Claude Opus 5.5 in most tests, though benchmarks show frontier models scoring similarly.
- Google emphasizes Carbon's improved safety features, with a 0% score in resisting prompt-injection attacks during the Grey Swan evaluation.
- Gemini 4 Carbon will have a staged release, mirroring Argon's rollout, through the Fairwinds cyber defense program via a voluntary testing framework.
- Initial access is expected for AI Ultra subscribers and developers using the Gemini API, with discounted rates of $2 per million input tokens and $10 per million output tokens, similar to Argon's introductory pricing.