Skip to content
MaungaAI news, explained

Models

Google keeps its most powerful AI for cybersecurity experts

Gemini 4 Carbon tops an independent ranking, and Google says it finds and fixes security flaws better than anyone else. But for now, only a select group of cyber defenders is using it.

2 min read8 sources, one officialWritten with AI · machine-translated

Share: WhatsApp · Email

AI-generated illustration for Maunga. Not a photograph.

The essentials

  • Google has unveiled Gemini 4 Carbon, an upgrade to Gemini 4 Argon, just two weeks after launching Argon (Ars Technica, The Verge).
  • For now, only a select group of cybersecurity defenders is using it. It will then reach partners in the Fairwind program and, with no date set, everyone else (MarkTechPost, The Verge).
  • It tops the Artificial Analysis index, which combines many tests, with 59 points, ahead of Claude Opus 5.5 and GPT-6 Astra (MarkTechPost, The Verge).
  • Claimed by one partyAccording to Google, Carbon writes code free of security flaws twice as often as Argon (MarkTechPost, Ars Technica).

Why it matters

Google is once again showing before opening up: the model exists, but the public can't use it. Giving it to defenders first suggests Google sees its ability to find and exploit flaws as sensitive. If its figures hold up, the people who protect systems will have a powerful tool before anyone else.

The details

According to Google, Carbon is designed for long, difficult programming jobs. Its internal teams watched it work for whole days on a single task, such as finding hidden flaws in huge programs and fixing them (MarkTechPost). In cybersecurity, Google pitches it for the full cycle: find the flaw, check whether it can be exploited and fix it. The Chrome team says its patches are ten times better than those from models outside Google (MarkTechPost, Ars Technica).

In the table Google publishes, Carbon comes first in 15 of 17 tests, although most rival figures are reported by each company itself (MarkTechPost). When exploiting real flaws (ExploitGym), it succeeds 54.8% of the time, compared with 42.4% for GPT-6 Astra. Google ran that test itself, with the system's protections turned off (MarkTechPost). When using a computer the way a person does (OSWorld-2.1), it scores 81.5%, below Claude Opus 5.5 (MarkTechPost).

What we don't know

  • When developers, businesses and users will be able to use it, and at what price: Google only says “as soon as possible.”
  • Whether the results hold up in independent tests: Google ran several of them itself, and most rival figures are reported by each company.
  • Whether it will be available in Europe when it opens up.

Background

On October 14, we reported that Google was opening Gemini 4 Argon, its previous top model, to paying subscribers and developers. Carbon arrives two weeks later and follows the same phased launch pattern (Ars Technica).

Sources

This story is based on 9 sources. In the text, each fact names in brackets who reports it.

Was this story useful? Yes Not really

Spanish original