Google's Gemini 4 Carbon Explained: The AI Model It Isn't Letting Everyone Use Yet

Google announced Gemini 4 Carbon on October 28, 2026, calling it its strongest AI model, with record cybersecurity scores. Here is what it can do, why only cyber defenders get it first, and what it will cost.

By Zeeshan Khalid · Updated 31 October 2026

Two weeks after launching Gemini 4 Argon, Google announced another model on October 28, 2026: Gemini 4 Carbon. Google calls it its strongest frontier model, and the unusual part is that most people can't use it yet.

What Gemini 4 Carbon is

Carbon is a general-purpose AI model that is better than Argon at coding, agent tasks, science and multimodal work. Google says its biggest improvement is cybersecurity: finding, exploiting and fixing software vulnerabilities.

Google also says Carbon is ranked first on the Artificial Analysis Intelligence Index (version 5.0), a popular combined leaderboard.

The cybersecurity numbers

Google published results on three security benchmarks:

Benchmark	What it tests	Carbon
SEC-bench Pro	Finding vulnerabilities	88.2%
ExploitGym	Exploiting vulnerabilities	54.8%
CWE-bench	Patching vulnerabilities	71%

Google also claims Carbon:

generates secure code at more than twice the success rate of Gemini 4 Argon, on Google's own tests,
performed "more than 10X better" than another lab's large model at producing high-quality patches.

Caveat: these are Google's own reported numbers. When it was announced, no independent evaluation had been published.

Why it isn't public yet

A model that is very good at finding security holes is just as useful to attackers as to defenders. So Google is doing a staged release:

First: a small group of cyber defenders (security teams) get access.
Then: wider access through Google's security programme.
Later: developers, businesses and consumers, with no date given.

Google says it is also going through the US government's voluntary review process before a wide release. Carbon uses Argon's safety systems plus extra work on honesty, prompt-injection resistance and "less overstepping" when it acts on its own.

What it will cost

Google has already listed general-availability API pricing:

$2 per million input tokens
$10 per million output tokens
95% discount on cached input

That is low for a top-ranked model, so price will not be what holds back developers once it opens up.

Why this matters

This is the first time a major AI lab has held back its strongest model mainly because of how good it is at cybersecurity. We will probably see more of this: the best models released first to defenders and then to everyone else.

For regular users, nothing changes today. Gemini 4 Argon is still the model in the Gemini app.

Prefer AI that never leaves your computer? See how to run local AI with LM Studio and Opera's new WebAI standard.