AI models

Gemini 4 Carbon, Google's new model: what it is, what it can do and when you can use it

On October 28, 2026, Google announced Gemini 4 Carbon, its new frontier model, about two weeks after opening up Gemini 4 Argon. Google calls it its strongest model, with state-of-the-art results in cybersecurity and long-horizon tasks. For now only a group of cyber defenders can use it. Here is what we know and what it means for you.

Published on 6 min readBy the editorial team of The AI Creator System

Leggi questo articolo in italiano

AI-generated cover: glowing carbon crystal lattice in cyan and violet on a midnight blue background with the headline Gemini 4 Carbon
Cover image generated with artificial intelligence. How we use AI

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google's new frontier AI model, announced on October 28, 2026, in a post by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. In the official announcement Google calls it its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities, meaning multi-step jobs that take hours or days.

Carbon follows Gemini 4 Argon, which Google opened to subscribers and developers on October 13, 2026; we covered it in Gemini 4 Argon: availability, pricing and how to use it. According to 9to5Google, Google attributes the fast follow-up to its researchers and models working together to push capabilities forward more quickly.

Google says its internal teams have seen Carbon work for days on complex jobs, such as root-causing and patching a bug in a large codebase, pushing through dead ends along the way. The company also reports gains in coding, agentic and scientific tasks, full-stack web development and multimodal understanding (handling text, images and other formats together).

Who can use Gemini 4 Carbon, and when it arrives

The short answer: almost nobody, for now. Google is following the same phased path it used for Argon:

  1. Starting this week: Carbon is available to a subset of cyber defenders.
  2. Next phase: access will expand to other trusted partners through the Fairwind Program.
  3. Developers, businesses and consumers: Google only says Carbon is "coming soon", without a date, a subscription plan or a list of countries.

Google adds that it is taking part in the U.S. government's voluntary pre-release model testing process. The announcement does not yet give an API model ID or confirm availability in the Gemini app, Google AI Studio or Vertex AI. For reference, Argon was announced on September 30, 2026, and reached Google AI Pro and Ultra subscribers on October 13, about two weeks later. That is a precedent, not a promise for Carbon.

How good is it? Google's reported benchmarks

All of the following numbers are reported by Google in its announcement and repeated by 9to5Google. Treat them as company figures until independent tests are available:

BenchmarkWhat it measuresReported result
Artificial Analysis Intelligence Index v5.0Overall capability index59, ranked first
SEC-bench ProFinding vulnerabilities88.2%, state of the art
ExploitGymExploiting vulnerabilities54.8%, state of the art
CWE-benchPatching vulnerabilities71.0%
WebDev ArenaWeb development+89 Elo, ranked fourth

Google notes that the cybersecurity benchmarks were run without guardrails. Compared with Argon, the company reports more than double the secure-code success rate on its internal benchmarks, and says the Chrome team saw more than 10x better high-quality patch generation than with another lab's large frontier model, which Google does not name.

More reliable than Argon? What Google says

For everyday AI users, the most interesting part of the announcement is reliability. Carbon uses the same multi-layered safety stack as Argon: model training, real-time misuse defenses, resilience to prompt injection (hidden instructions in pages or documents that try to hijack the AI) and monitoring of the model's chain of thought.

In Google's internal testing, compared with Argon:

  • unsupported claims are cut by roughly half;
  • cases where the model does more than asked fell by 26%, and cases where it does less than asked fell by 29%;
  • high-severity misalignment alerts in live internal coding environments dropped by almost 80%;
  • on the Gray Swan indirect prompt injection benchmark the attack success rate is 0.0%, versus 0.7% for Argon;
  • on Terminal-Bench 4.0 cheating fell to 0.0%, with cheating attempts at 0.4%.
Worth knowing. Fewer made-up claims does not mean zero mistakes. With Carbon too, check facts, names and quotes before you publish them.

Gemini 4 Carbon API pricing

Google has already published pricing, in US dollars per million tokens (a token is a chunk of text, on average a little less than one word):

ItemIntroductory priceAfter the introductory period
Input$2$4
Output$10$20
Cached input95% off the input pricenot stated

The introductory price matches Argon's. The Carbon announcement does not say when the introductory period ends. As a rough example, a request that sends 100,000 tokens and gets 20,000 back would cost $0.20 + $0.20, about 40 cents at the introductory price and about 80 cents afterwards.

What it means for you

If you create content or run a business, Carbon does not change your work today, because you cannot use it yet. It does show where things are heading. Late-2026 frontier models are betting on long, autonomous tasks and fewer made-up answers, as we also saw with OpenAI's GPT-6.1 Terra. Here is how to get ready:

  1. Write down your procedures. An agent that works for hours performs best with clear instructions: list step by step how you build a content calendar, a quote or a report. You can already use them with Argon or Claude, and Claude Business OS shows you how to turn those procedures into a working setup with ready-made projects and instructions.
  2. Prepare your tests. Pick three real tasks and save the answers from the model you use today. When Carbon arrives, give it the same tasks and compare the results on your own work, not on leaderboards.
  3. Automate one process at a time. Ready-made prompts, workflows and templates for marketing, sales and customer service are in the AI Business Toolkit: they work with any language model and will carry over to Carbon.
  4. Watch your costs. If you build on the API, prices double after the introductory period, so measure the cost per request from your first tests.

Meanwhile, Google is also pushing into video and interactive worlds: read our piece on Google DeepMind's Genie 4 and our Gemini Omni guide. If you want a complete path to producing professional content with AI, start with The AI Creator System.

Frequently asked questions

What is the Gemini 4 Carbon release date for everyone?

Google has not given a date. Since October 28, 2026, Carbon has been available to a group of cyber defenders, and it will then reach other trusted partners through the Fairwind Program. For developers, businesses and consumers Google only says "coming soon".

What is the difference between Gemini 4 Carbon and Gemini 4 Argon?

Carbon is the successor to Argon and, according to Google, is stronger at coding, agentic tasks, cybersecurity and long-horizon work. Google also reports about half as many unsupported claims as Argon. The introductory API price is the same.

How much does Gemini 4 Carbon cost?

Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period it rises to $4 and $20.

Will Gemini 4 Carbon be in the Gemini app?

Google says Carbon is coming soon to consumers, but its October 28, 2026 announcement does not name the app or subscription plans. Argon, the previous model, reached the Gemini app for Google AI Pro and Ultra subscribers.

Read next

Sources checked on October 29, 2026