Gemini 4 Carbon Tops Index, Delays Developer Access

This is a rational safety policy wrapped in convenient scarcity marketing: the cyber-risk argument is credible, but “#1 model you can’t use” also maximizes hype.

• –Artificial Analysis confirms the lead, but Carbon is not dominant across every coding or knowledge-work evaluation.
• –Google’s cybersecurity benchmarks were run without guardrails, so production performance may differ materially.
• –Anthropic and OpenAI are pursuing similar defender-first programs, suggesting a genuine industry-wide shift around dual-use cyber capabilities.
• –Developers cannot yet validate Carbon’s API behavior, reliability, token usage, or real-world cost.
• –The model’s introductory pricing is attractive, but its eventual $4 input and $20 output rates make task-level economics more important than leaderboard rank.