---
title: "Gemini 4 Carbon: Google unveils its most powerful model, a cybersecurity specialist, but isn't letting it out yet"
url: https://hernandez.systems/en/articles/gemini-4-carbon-google-unveils-its-most-powerful-model-a-cybersecurity-specialist-but-isnt-letting
lang: en
type: NewsArticle
section: AI
published: "2026-10-30T03:31:02+00:00"
modified: "2026-10-30T06:55:06+00:00"
publisher: Osky Networks
image: https://oskydj.sfo3.cdn.digitaloceanspaces.com/calamardo/news_item/2026-10-30/e11-05a105c9.jpg
alternate_es: https://hernandez.systems/es/articulos/gemini-4-carbon-google-estrena-su-modelo-mas-potente-especialista-en
---

# Gemini 4 Carbon: Google unveils its most powerful model, a cybersecurity specialist, but isn't letting it out yet

> Google introduced Gemini 4 Carbon on October 28, its new frontier model, barely two weeks after launching Gemini 4 Argon. The company describes it as its strongest model, with cybersecurity and long-task capabilities that, by its own numbers, are at the highest level ever measured. The detail that matters most if you write code: you can't use it yet. Google is releasing it first to a group of cybersecurity defenders and gave no date for developers, businesses, or consumers.

Google introduced Gemini 4 Carbon on October 28, its new frontier model, barely two weeks after launching Gemini 4 Argon. The company describes it as its strongest model, with cybersecurity and long-task capabilities that, by its own numbers, are at the highest level ever measured. The detail that matters most if you write code: **you can't use it yet**. Google is releasing it first to a group of cybersecurity defenders and gave no date for developers, businesses, or consumers.

## What Google announced

The announcement is signed by Koray Kavukcuoglu, senior vice president at Google DeepMind and Google's chief AI architect. Here are the main points:

- **Number 1 on the Artificial Analysis Intelligence Index v5.0**, with 59 points.
- **Cybersecurity:** 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (checking whether they can be exploited), and 71.0% on Collinear's CWE-bench (patching them at the root cause). Google clarifies that these tests were run **with safeguards turned off**.
- **Long tasks:** according to Google, its teams have seen Carbon work for days on end on multi-step problems, such as finding and fixing subtle bugs in large codebases and low-level systems.
- **Web development:** up 89 Elo points, landing in fourth place on WebDev Arena with 1768.
- **Secure code:** in internal tests, Carbon doubles Argon's success rate at generating secure code. Google also cites the Chrome team, which reported that Carbon was more than 10 times better than "another lab's large frontier model" at generating quality patches. It doesn't say which one.

## How much it will cost

Carbon will launch at the same introductory price as Argon: **2 dollars per million input tokens and 10 per million output tokens**, with a 95% discount on cached input tokens. When the introductory period ends, the price goes up to **4 and 20 dollars**. Google doesn't say when that period ends, and Artificial Analysis confirms there's no end date.

According to Artificial Analysis, the model has a one-million-token context window, accepts text, image, video, and voice as input, and only responds in text.

## Who can use it and when

Here's the fine print. Google says Carbon is part of the U.S. government's voluntary pre-release testing process. Starting this week it's opening it to "a subset of cyber defenders" and will then add trusted partners through its **Fairwind Program**. General availability will come "as soon as possible," after gathering feedback and fine-tuning the safeguards.

Fairwind is DeepMind's program that gives priority defenders, such as governments, healthcare providers, and telecoms, early access to advanced models. The page says it currently works with more than 650 partners worldwide and vets every applicant. If your organization doesn't qualify, Google suggests using CodeMender, its tool for finding and patching vulnerabilities, with the public models.

## What independent measurement says

Artificial Analysis published its own evaluation the same day, and it's worth reading because it adds nuance to the announcement. It confirms first place on the overall index with 59 points and that Carbon earns the top score on 4 of the index's 9 tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. It also makes clear that it **doesn't lead in coding or knowledge work**.

| Metric (Artificial Analysis) | Gemini 4 Carbon | Reference |
|---|---|---|
| Coding Agent Index | 59 (with Antigravity CLI) | Claude Opus 5.5 in Claude Code: 62 |
| FrontierCode v1.1 | 51% (2nd place) | Claude Opus 5.5: 56% |
| Terminal-Bench 4.0 | 64% (2nd place) | — |
| AA-Briefcase (Elo) | 1609 | Argon: 1483 · Kimi K3.1 Pro: 1678 |
| Cost per index task | 7.76 dollars | Opus 5.5: 10.04 · GPT-6 Astra: 6.25 |
| Hallucination rate | 11% | Argon: 4% · Opus 5.5 and GPT-6 Astra: 50% |

In agentic coding, Artificial Analysis puts it behind only Claude Opus 5.5, and stresses that it's a notable improvement because Google had been lagging in that area. On cost, at standard pricing, each index task comes out about 23% cheaper than with Opus 5.5 and uses 29% fewer output tokens (123,000 versus 173,000), although it's still more expensive than GPT-6 Astra.

There's one number worth keeping in mind: the hallucination rate goes up from 4% on Argon to 11% on Carbon. It's still the second lowest of all the models Artificial Analysis measures, but it's almost triple that of its predecessor.

![Gráfica de barras con tres comparativas de Artificial Analysis: Coding Agent Index (Gemini 4 Carbon 59, Claude Opus 5.5 62), costo por tarea del índice (Carbon 7.76 dólares, Opus 5.5 10.04, GPT-6 Astra 6.25) y tasa de alucinación (Argon 4%, Carbon 11%, Opus 5.5 y GPT-6 Astra 50%)](https://oskydj.sfo3.cdn.digitaloceanspaces.com/calamardo/news_item/2026-10-30/e11-comparativa-84a63d93.png)

## What Google says about the model's safety

Besides going after other people's vulnerabilities, Google boasts that Carbon behaves better. These figures come from its internal tests and nobody outside has verified them yet:

- On Terminal-Bench 4.0, successful cheating dropped to 0.0% and cheating attempts to 0.4%.
- In real programming environments with its teams, severe misalignment alerts dropped by almost 80%.
- Unsupported claims are cut roughly in half compared to Argon.
- It oversteps 26% less on what it's asked to do and leaves 29% fewer tasks half-finished.
- On Gray Swan's indirect prompt injection benchmark, the successful attack rate went from 0.7% with Argon to 0.0%.

## My take

The most interesting part isn't first place on the index, which at this pace lasts a few weeks. It's the shape of the launch. Google released a model whose main selling point is that it **finds and exploits vulnerabilities better than anyone**, and for that very reason it isn't releasing it to the public. The cybersecurity numbers were measured without safeguards, in other words, it's exactly the kind of capability you don't want in just anyone's hands. Hipertextual ties the launch to the recent wave of attacks carried out by autonomous agents against companies and governments, although Google doesn't mention it in its announcement.

For developers, three things change in practice. First, there's nothing to integrate yet: if your stack depends on Gemini, you stay on Argon. Second, the price is already announced and it's the same as Argon's, so switching from one to the other shouldn't move your budget, although it's wise to prepare for the jump to 4 and 20 dollars when the introductory period ends. Third, if you work in security at a large organization, Fairwind is the only way in for now.

And a healthy warning: almost all the cybersecurity and alignment figures come from Google. The only independent measurement available today, Artificial Analysis's, confirms the overall lead, but also shows that in pure coding the throne still belongs to someone else. Until there's open access and third-party testing, Carbon is a very well-documented promise, not a tool you can use.

## Sources

- [Gemini 4 Carbon: our next frontier model (Google)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [Announcement on the Google DeepMind blog](https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/)
- [Fairwind Program (Google DeepMind)](https://deepmind.google/fairwind-program/)
- [Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5 (Artificial Analysis)](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1)
- [Google lanza Gemini 4 Carbon, una IA que puede detectar fallos antes que los hackers (Hipertextual)](https://hipertextual.com/inteligencia-artificial/gemini-4-carbon-google-ciberseguridad-ia/)