Google announced Gemini 4 Carbon on October 28, 2026, describing it as its strongest model, with cybersecurity and long horizon capabilities. It is not an open launch. For now only a group of cybersecurity defenders uses it through the Fairwind program, and there is no date for the general public. As an artificial intelligence agency, we follow these launches because they end up affecting the tools businesses use. In this article we explain what it is, who can use it, what the results measure, and what is worth doing now.
What is Gemini 4 Carbon and when was it announced?
Gemini 4 Carbon is Google's new frontier model, announced on October 28, 2026. The Google AI post presents it as its strongest model, with state of the art cybersecurity and long horizon capabilities. The latter means working on complex tasks over extended periods. According to Constellation Research, it arrives about two weeks after Gemini 4 Argon, which suggests a faster release pace.
Google states that Carbon is a step up across most benchmarks and that it stands out in autonomous coding and cyber defense. It also says it can work on complex tasks for days without taking shortcuts. It has not published the evaluation method for that claim or an independent validation, according to Hora de Codar.
It is worth clarifying what it is and what it is not. Carbon is not a new Google Ads or Search feature. It is a language model, and its relevance for marketing is indirect for now.
Who can use Gemini 4 Carbon today?
Only a restricted group of cybersecurity defenders. Google offers it first through Fairwind, its program for defenders, and plans to expand access gradually to trusted partners. According to Tom's Hardware, developer access will come only when Google refines the model's cybersecurity safeguards.
Business Model Analyst adds data about Fairwind. According to that outlet, the program launched on September 2, 2026 with more than 650 partners and targets priority defenders, such as governments, healthcare providers, and telecom operators. Participants can only do defensive work and may not share, redistribute, or sell access. To request it, organizations must contact their Google Cloud account team.
The full eligibility criteria are not published. There is also not yet a public API, a model identifier for API calls, or access in the Gemini app or in AI Mode in Search, according to Hora de Codar. That source warns that any tutorial or model identifier circulating before official confirmation should be treated as unverified.
What do the cybersecurity results say?
Google reports three cybersecurity benchmarks, according to Hora de Codar. On SEC-bench Pro, which measures finding vulnerabilities, Carbon scores 88.2%. On ExploitGym, which measures exploitation, 54.8%. And on CWE-bench, which measures fixing flaws, 71.0%.
There is a key warning. According to the announcement, these tests were run without safeguards. That means the results may not reflect the product that is eventually released. They are also Google's own figures, not a third party's.
This capability has two readings. For a defender, a model that finds and fixes flaws speeds up the work. But the same skill, in the wrong hands, could speed up an attack. That tension explains why Google restricts access and why it offers Carbon first to verified defenders.
Is Gemini 4 Carbon really the best model?
In an independent ranking, it is first by a small margin, but it does not win at everything. Artificial Analysis gives it 59 points on its Intelligence Index V5, compared with 58 for Claude Opus 5.5 and 57 for GPT-6 Astra, according to Tom's Hardware. Carbon leads four of the nine evaluations. That same firm indicates that it trails Claude Code with Opus 5.5 in programming and trails other models in knowledge work.
There is a data point that qualifies Google's narrative. Artificial Analysis measures a hallucination rate of 11% for Carbon, compared with 4% for Gemini 4 Argon, although the model answers more questions correctly, 58% compared with 48%. Google, for its part, states that Carbon cuts unsupported claims twice as much as Argon and that it declines actions it cannot complete. The two data points do not necessarily contradict each other, because they measure different things, but it is worth keeping both in view.
On cost, Artificial Analysis calculates 7.76 dollars per index task for Carbon, compared with 10.04 for Claude Opus 5.5 and 6.25 for GPT-6 Astra. These are rankings from a specific date and may change with each model update.
How much will it cost and why is access so restricted?
When it reaches the API, Google says Carbon will cost 2 dollars per million input tokens and 10 per million output tokens, with a 95% discount on cached input. It is an introductory price. Afterward it will rise to 4 and 20 dollars. Google has not said when the introductory period ends, and Business Model Analyst notes that it has not published a price for the defender version either.
On the reason for the restriction, Google cites its voluntary pre release tests with the United States government and the work of refining safeguards. Business Model Analyst offers another reading: the requirement to go through the Google Cloud account team would also work as a sales funnel. It is the author's opinion, and the author acknowledges that this channel is an efficient way to vet applicants. What is documented is that access goes through Google Cloud.
What changes for a business or a marketing agency?
Almost nothing in the short term, and it is best not to change tools because of this announcement. Carbon is not available to the public or in the API, so you cannot use it today in your workflows. These are recommendations from our artificial intelligence agency, not Google instructions.
First, do not trust tutorials or model identifiers that promise access. Second, when the API arrives, test the model with your own tasks before migrating, because general rankings do not say how it performs in your case. Third, measure quality, cost, and error rate with your data, since the hallucination figure calls for caution. Fourth, if your business depends on a website, review the security basics: updates to your content manager and plugins, backups, and user access. Fifth, ask your vendors which models they use and what data they send to them.
Consider a typical case, not real data but an example of how it might look. A dental clinic in Dallas has a WordPress site managed by its agency. Its team notes that Carbon is not available, so it changes nothing. Instead, it takes the chance to review which plugins are outdated, check its backups, and remove accounts of users who no longer work there.
When does it make sense to work with an artificial intelligence agency?
When you want to use AI in your business, but every week a new model appears and you do not know which deserves a test. An artificial intelligence agency can filter vendor announcements, design tests with your real tasks, and measure quality, cost, and errors before changing anything. It can also organize what data is sent to each tool. If you want us to review how to use AI in your company, you can contact the JP Director team.
This topic connects with others we have published. To see how other frontier models compare, read our analysis of GPT-6 Astra and Fable 5.1. And to understand what permissions to give an AI agent, review our blog on agent permissions at Microsoft.
Frequently Asked Questions
What is Gemini 4 Carbon?
It is Google's strongest frontier model, announced on October 28, 2026, with cybersecurity and long horizon capabilities. For now only cybersecurity defenders use it through Fairwind.
Can I use Gemini 4 Carbon today?
Probably not. There is no public API or access in the Gemini app or in AI Mode, and Google has given no date for the general public. Initial access is for verified defenders.
What is the Fairwind program?
It is Google's program that gives verified defenders access to its cybersecurity models, such as governments, healthcare providers, and telecom operators. Use is limited to defensive work and the full criteria are not published.
Is it better than Claude Opus 5.5 and GPT-6 Astra?
On Artificial Analysis's Intelligence Index, Carbon scores 59 against 58 and 57, and it trails Claude Code with Opus 5.5 in programming. It is a small lead on a specific date, and it does not replace a test with your own tasks.
Last updated: November 4, 2026. This article will be updated when Google confirms API access, Fairwind criteria, or the final price of Gemini 4 Carbon.
