Google announces Gemini 4 Carbon, its latest frontier AI model that you can't use

Key Points:

• Google has announced Gemini 4 Carbon, an improved version of its Gemini 4 Argon AI model, focusing on enhanced coding, agentic tasks, and scientific analysis capabilities, with a staged release planned.
• Gemini 4 Carbon shows significant advancements in cybersecurity, including better vulnerability detection, validation, and remediation, and is twice as likely to generate vulnerability-free code compared to Argon.
• The model outperforms competitors like GPT-6 Astra and Claude variants in benchmarks, though the margin of improvement is relatively small, and it achieved a perfect score in resisting prompt-injection attacks in the Grey Swan evaluation.
• Google is taking a cautious rollout approach, complying with US government voluntary testing frameworks, and plans to release Carbon initially to AI Ultra subscribers and developers via the Gemini API with discounted introductory pricing.
• Pricing for the Gemini API will start at $2 per million input tokens and $10 per million output tokens during the introductory period, with no specified end date for the discount, after which the rates will double.