Gemini 4 Carbon leads benchmarks, without public access

Google presents Gemini 4 Carbon, an AI model designed for cybersecurity and long coding sessions. It tops several rankings but remains reserved for a few cybersecurity defenders for now.

Google has unveiled Gemini 4 Carbon, an AI model designed for cybersecurity and long-term coding work, leading 15 out of 17 benchmarks according to the company's charts. The downside is clear: the public does not yet have access, with only a few cyber defenders able to try it initially. In a market where OpenAI and Anthropic are advancing quickly, this return to the top may only last a moment.

A return to the top, but on a still closed stage

With Carbon, Google shows clear progress after a 2026 spent chasing its rivals. The model scores 59 on the Artificial Analysis Intelligence Index, placing it ahead of Claude Opus 5.5, with a 9-point gain compared to Argon launched four weeks earlier. Its most marked differences come from reasoning over long contexts (+24 points), hacking (+12), workflow automation (+11), and mathematics (+8).

The problem is that these results remain mostly theoretical for the general public for now. Google has given no date for a wider rollout, and Argon had already followed a similar logic, with limited access initially granted to cyber defenders. The demonstration is solid on paper but impossible to test freely. The medal is shiny, the entry door remains closed.

Why price matters as much as ranking

Carbon also arrives with an API pricing of 2 dollars at entry and 10 dollars at exit per million tokens, about half the price of Opus 5.5 and a fifth of Astra at launch, according to the source. After the discount phase, the price rises back to Opus's level. This detail is not trivial: in real use, cost can weigh as much as raw performance, especially for teams industrializing agents or security tools.

In this context, Google is clearly trying to show it can return to the race not only with a high-performing model but also with a more aggressive distribution strategy. The schedule remains crucial. If competitors publish at the same speed, the benchmark crown could change hands before Carbon even leaves its reserved showcase.

What Google is trying to demonstrate with Gemini 4 Carbon

The launch sends a signal on two levels. First, Google shows it can align two major models within weeks, after a period where the company seemed behind on cutting-edge AI. Second, it explicitly targets uses where reliability and persistence matter more than a simple quick exchange: cybersecurity, long code, automation. This is where part of companies' spending happens, not just in demonstrations that shine on keynotes.

For technical decision-makers, the message is clear: benchmarks are becoming favorable to Google again, but actual availability remains the true measure. Without broader access, it is impossible to know if Carbon keeps its promises in real environments, with production, compliance, and supervision constraints.

Key points

  • Gemini 4 Carbon dominates 15 benchmarks out of 17 according to Google.
  • Score of 59 on the Artificial Analysis Intelligence Index.
  • Google has not announced a public access date.
  • Initial access reserved for certain cyber defenders.
  • API entry price: 2 dollars per million tokens.
  • Claude Opus 5.5 and Astra remain the main cited benchmarks.

By the numbers

  • 15 out of 17 — benchmarks dominated by Gemini 4 Carbon, according to Google.
  • 59 — Carbon's score on the Artificial Analysis Intelligence Index.
  • 9 points — gain shown compared to Argon, launched four weeks earlier.
  • 2 / 10 dollars — launch API price per million tokens, entry/exit.
  • 15 points — maximum gap mentioned on Carbon's gains compared to other models.

To read