Gemini 4 Carbon tops AI benchmark at 59, but only cyber defenders get it first

Gemini 4 Carbon's first-place benchmark score comes with a caveat that matters more than the ranking itself: the model is not yet broadly available. Google DeepMind's plan is to start with a subset of cyber-defense users and expand from there to trusted partners, then enterprises and the general public "as soon as possible" — a phrase that carries no date. Ars Technica highlighted the gap directly, calling it in its headline the latest model you cannot use.

The practical alternative today is Argon, released two weeks earlier, which is the model currently available to paying customers. Carbon's pricing is already fixed at the same level as Argon, $2 per 1 million input tokens, suggesting Google intends the two to sit side by side once access opens.

A further wrinkle: the composite benchmark that produced the 59-point score has recently changed how it is calculated, and questions about that methodology have appeared on overseas forums. That does not erase the security results — 88.2% on finding flaws and 71.0% on fixing them — but it does mean the headline "first place" deserves to be read alongside the access restrictions and the scoring debate.