Google reserves its most powerful AI, Gemini 4 Carbon, for cyber defense experts first

Two weeks. That is the gap between the opening of Gemini 4 Argon to developers and the announcement of its successor. On October 28, Koray Kavukcuoglu, a senior executive at Google DeepMind, unveiled Gemini 4 Carbon, which Google describes as its most powerful model. It is designed for long engineering tasks and for cybersecurity.

There is one big catch: almost nobody can use it. This week, only a small group of cyber defense professionals has access, through a Google program called Fairwind. The company promises to open it to developers, businesses and the general public "as soon as possible," with no date. It says it first wants to gather feedback and refine the model's safeguards. Nothing has been said about availability in Europe.

First on an independent index, by one point

Artificial Analysis, an independent evaluation firm, ranks Carbon first on its Intelligence Index (version 5.0). It scores 59 points, against 58 for Anthropic's Claude Opus 5.5, 57 for OpenAI's GPT-6 Astra and 50 for Argon, according to VentureBeat. The lead over its rivals is therefore minimal. The index was also overhauled on October 27, so its scores cannot be compared with the old ones.

The picture is more nuanced in the details. On the same evaluator's coding agent index, Carbon scores 59, behind Claude Code with Opus 5.5 (62). It answers knowledge questions correctly more often than Argon. But its hallucination rate, as measured by Artificial Analysis, rises from 4% to 11%: when it does not know, it is more likely to make up an answer. At Vals AI, another evaluator, it ranks second out of 29 models, behind Claude Sonnet 5.5.

Security scores measured by Google itself

The most spectacular results come from Google itself. The company reports 88.2% on SEC-bench Pro (vulnerability discovery) and 54.8% on ExploitGym (vulnerability exploitation). It says it doubled the success rate in secure coding compared with Argon. It also claims to have observed no successful prompt injection attacks on a test from the company Gray Swan, compared with 0.7% for Argon. These figures have not been independently verified. Google specifies that its cybersecurity evaluations were run "without safeguards."

What this changes for a European business

In the short term, not much. Argon remains the model you can use today, and its launch pricing ends on November 30. Google says Carbon's launch price is the same as Argon's: $2 per million tokens (units of text) for input and $10 for output. It will then rise to $4 and $20, with no switch date announced.

• Do not build a project on Carbon until a public API and a date are announced.
• Compare on your own use cases: a one-point lead on a general index says nothing about your line of work.
• If the reliability of answers matters to you (legal, customer support), watch the rise in hallucinations measured by Artificial Analysis.

By reserving its best model for defenders, Google is repeating the approach it already used for Argon, which was first opened through Fairwind on September 30. It remains to be seen how long this restricted period will last, and whether Europe will be served at the same time as the United States.