Skip to content
Advertisement
Google’s Gemini 4 Carbon AI model goes to cybersecurity defenders first

Google’s Gemini 4 Carbon AI model goes to cybersecurity defenders first

Photo: Tima Miroshnichenko / Pexels

Alphabet CEO Sundar Pichai says the security-focused model will roll out gradually through the Fairwind Program before reaching a wider audience

Alphabet CEO Sundar Pichai has a new AI model to show off. Most people will have to wait in line to use it.

Pichai announced that Gemini 4 Carbon, Google’s latest model, will first roll out to cybersecurity partners. The model is built with a heavy emphasis on finding software flaws, testing whether they can be exploited, and fixing them.

What Carbon is built to do

The launch came on October 28, 2026. Gemini 4 Carbon targets three core security tasks: vulnerability discovery, exploitation validation, and software patching. It also comes with strong coding capabilities.

The benchmark numbers back up the pitch. Carbon scored 59 on version 5.0 of the Artificial Analysis Intelligence Index, the top score on that ranking.

On security-specific tests, the results looked like this:

  • SEC-bench (vulnerability finding): 88.2%
  • ExploitGym (exploitation): 54.8%
  • CWE-bench (patching): 71.0%

All three figures mark a significant improvement over Gemini 4 Argon, the model Carbon replaces. On vulnerability discovery specifically, Carbon’s success rate more than doubles Argon’s.

Advertisement

Pricing, meanwhile, stays put. Carbon costs $2 per million input tokens and $10 per million output tokens, the same rates Google charges for Argon.

Two weeks, two models

Carbon arrived just two weeks after Gemini 4 Argon. The quick turnaround reflects Google’s use of AI-assisted model development.

Initial access begins this week for trusted cybersecurity defenders through Google’s Fairwind Program. The program exists to manage a responsible rollout and keep testing safe.

Pichai stressed the importance of rolling Carbon out responsibly, with strengthened safeguards in place. Google plans a phased expansion that would bring the model to more developers, businesses, and eventually consumers.

Investors seemed to like what they heard. Alphabet shares rose approximately 1.5% after the announcement.

Why the cautious rollout matters

The gated launch makes more sense once you look at what Carbon is good at. A tool that excels at finding vulnerabilities and confirming they can be exploited is useful to defenders. The same skills would be useful to attackers.

Starting with vetted security teams gives Google a chance to see how the model performs in real defensive work. The Fairwind Program effectively functions as a controlled test environment with real stakes.

The exploitation score is the number worth watching. At 54.8% on ExploitGym, Carbon is the weakest of the three security benchmarks by a wide margin.

What this means

Gemini 4 Carbon is positioned against OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5.

By leading with security, Google is staking a claim on a specific, high-value niche rather than just general-purpose intelligence. Holding pricing flat while improving performance also puts pressure on rivals.

Google has not tied the broader expansion to specific dates. Developers and businesses eager to try Carbon will be watching for the next stage of access.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.