Key takeaways
- Google announced Gemini 4 Carbon on Wednesday, October 28, 2026, but this week only a select group of cyber defenders can use it.
- The next stage runs through the Fairwind Program, which Google says gives governments, critical infrastructure operators and major tech platforms early access, with more than 650 participating organizations.
- Google ran its security benchmarks with no guardrails, and no source reports independent verification, so public-release performance may differ.
- Google gave no date for developers, businesses or consumers, only 'as soon as possible', and hasn't said who counts as a cyber defender.
Google announced Gemini 4 Carbon on Wednesday, October 28, 2026, and most people can’t touch it yet. This week it goes to a select group of cyber defenders. After that it widens in stages to other trusted partners through Google’s Fairwind Program, and only then to developers, businesses and consumers.
If you run security for an organization, the practical question isn’t how good the benchmarks look. It’s whether you’re in line, and where.
Who can use Gemini 4 Carbon right now?
A select group of cyber defenders, starting this week. Google hasn’t said how that group was chosen or what makes someone a defender for this purpose. Everyone else waits for later stages that have no dates attached.
The order, as Google describes it: the initial defender group, then other trusted partners through Fairwind, then Google’s product surfaces for developers, businesses and consumers. Google says that last step happens “as soon as possible,” after it gathers feedback and iterates on guardrails. That’s the whole timeline on offer.
Pricing exists already. Launch pricing runs $2 per million tokens sent in and $10 per million tokens generated, rising to $4 and $20 when the introductory period ends. Cached input tokens get 95% off the input price. Treat those as launch numbers for the initial access group, not a general-availability price list. Google hasn’t said how long the introductory period runs.
How the Fairwind Program works
Fairwind lets governments, critical infrastructure operators and large tech platforms try advanced cyber-defense models ahead of the public. Google says more than 650 organizations take part. Beyond those categories, the membership list and the criteria for joining haven’t been published in the coverage so far.
That leaves a gap for anyone sizing up their chances. A national agency or a large platform fits the description. A regional hospital system or a mid-size SaaS company might, as a critical infrastructure operator or not, and nothing reported says how to apply. If you think you belong, the sensible move is to ask your Google account contact directly rather than wait for a public form.
Fairwind isn’t new with Carbon. Some partners already have exclusive access to the cybersecurity defense capabilities of Gemini 4 Argon, the model Google released before this one.
Argon already ran this playbook
Argon followed the same limited-access plan and has since reached the Gemini app and AI search mode, according to Engadget. So the pattern has a precedent: defenders and partners first, broad release afterward.
How fast that happened is disputed. The Hacker News says Argon launched less than a week before Carbon. SecurityWeek, Neowin, 9to5Google and Engadget say two weeks, and PCMag says only that it came earlier in the month. Pick any of those and the cadence is quick. Google credits its researchers and models working together to move faster, and says Carbon was improved using feedback on Argon.
Quick doesn’t mean Carbon follows the same schedule. Google gave no date, and Argon’s gap isn’t a promise.
Why labs gate their strongest security models
The capability that helps a defender is the same one that would help an attacker. Google describes exploitation testing as a defensive step: defenders use models to check which vulnerabilities are actually exploitable so they can decide what to fix first. I’d guess that dual use is the reason for the staged release, though Google’s stated reason is about gathering feedback and tuning guardrails.
There’s a catch. Google says it ran its security benchmarks with no guardrails switched on. The model that reaches the public will have restrictions the tested one didn’t, and what those cost in capability isn’t known.
Google’s scores, as quoted by DeepMind’s Koray Kavukcuoglu, are specific:
“landing a firm lead in SEC-bench Pro (88.2%) for finding vulnerabilities, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching”
— Koray Kavukcuoglu, Google DeepMind, via The Hacker News
In Google’s comparison, Carbon beat OpenAI’s GPT-6 Astra, along with Anthropic’s Claude Fable 5.1 and Opus 5.5, on all three. All of it is Google’s own testing. Google’s Chrome team says Carbon’s patches were more than 10 times better than a rival lab’s large model, which it doesn’t name. Wiz, which Google owns, says it saw Carbon top web penetration testing benchmarks. Neither is independent, and no source reports outside verification.
Google also says it’s taking part in the US government’s voluntary process for testing models before release. What that testing involves, and whether Carbon has cleared it, hasn’t been reported.
Other labs are treating cyber capability differently too
The Hacker News reports that Anthropic disclosed this week that an unreleased internal research model found critical vulnerabilities in widely used software. They included a one-click remote code execution flaw in Chrome, a one-click exploit chain in Safari, an Android lock screen bypass, and a Windows HTTP.sys denial-of-service flaw (CVE-2026-33096, CVSS 7.5). Google, Apple and Microsoft have fixed all of them.
The same outlet says OpenAI released GPT-6.1 Terra and that OpenAI says it meets the Critical cybersecurity threshold under its Preparedness Framework. Neither was part of Google’s comparison. For a side-by-side of the broader field, see our scorecard of Meta Muse Wonder, Opus, Astra and Gemini 4, and for the OpenAI side, our guide to GPT-6.1 Terra versus Luna, Sol and Astra.
What this means if you run security
Don’t plan around Carbon access you don’t have. If you’re in the initial group or a Fairwind partner, you’re evaluating a model whose tested performance had no guardrails, so benchmark it on your own code and your own triage workflow before you trust any number from Google.
If you’re outside, the nearer-term effect is on the patch queue. Anthropic’s disclosure shows an AI model finding browser, mobile and Windows flaws that vendors then fixed. Expect more discovery at more vendors and keep patch cycles short. Our note on the latest iOS security fixes shows what that looks like on the consumer side.
Accounts are the other lever. Admin and cloud console logins are worth protecting with a hardware security key such as a YubiKey 5 NFC, which doesn’t depend on how fast any model finds the next bug.
For budgeting, wait. Launch pricing will rise, general availability has no date, and the guardrailed model may behave differently from the one Google benchmarked.
Open questions on access and testing
Several things that matter for deciding timing are still unanswered:
- Who qualifies as a cyber defender? Google hasn’t defined it, and hasn’t said how it will prevent misuse of exploit-generation capabilities.
- What does the government testing cover? Google says it takes part in the voluntary pre-release process. Whether Carbon has cleared it isn’t reported.
- What will the public guardrails change? The benchmarks were run without them.
- How long is the introductory price period? Pricing rises to $4 and $20 afterward, with no date given.
- Will outsiders reproduce the results? Every benchmark figure so far comes from Google or a Google-owned company.
- How does Carbon compare with GPT-6.1 Terra and Anthropic’s unreleased model? Neither was in Google’s comparison.
Tools & products mentioned
- YubiKey 5 NFC on Amazon — Check current price
Disclosure: some links in this article are affiliate links or our editor's own courses and book. If you buy through them, TechyCluster or its editor may earn money at no extra cost to you. This does not influence what we cover or how we cover it. As an Amazon Associate, TechyCluster earns from qualifying purchases.
Frequently asked questions
Can I use Gemini 4 Carbon today?
Only if you're in the select group of cyber defenders Google opened access to this week. Broader access to trusted partners, then developers, businesses and consumers, is planned with no dates beyond 'as soon as possible.'
What is the Google Fairwind Program?
Google says Fairwind lets governments, critical infrastructure operators and major tech platforms try advanced cyber-defense models early. More than 650 organizations participate, and some already have exclusive access to Argon's cyber defense capabilities.
How much does Gemini 4 Carbon cost?
Introductory pricing is $2 per million tokens in and $10 per million tokens out, rising to $4 and $20 later. That applies to the initial access group, not a general-availability price.
Are Google's Carbon benchmark results independent?
No. They're Google's own comparisons or internal tests, and Google says it ran the security tests without guardrails. No source reports independent verification yet.
Get the next one in your inbox. One email a day with the tech stories that matter, explained in plain English. Subscribe free.
Sources
This article was compiled from reporting by the following outlets. Links go to the original reports.
- The Hacker News: Google's Gemini 4 Carbon Leads Benchmarks for Finding, Exploiting, and Patching Vulnerabilities
- SecurityWeek: Google Touts Powerful Cybersecurity Capabilities in Gemini 4 Carbon Model
- Neowin: Google Gemini 4 Carbon dethrones GPT-6 Astra and Claude Opus 5.5 at a fraction of the price
- PCMag: With Gemini 4 Carbon, Google Touts Big Leap in Cybersecurity Capabilities
- 9to5Google: Google announces Gemini 4 Carbon with ‘significant gains across coding and agentic tasks’
- Engadget: Google's Gemini 4 Carbon promises 'significant leap' in AI coding and cybersecurity
