Introduction: the model launch I could not try
I have a habit with big AI launches: I open the announcement, skim the benchmarks, and then go try the thing myself before I write a word about it. This week that habit broke. On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, which Google calls its strongest model yet for cybersecurity and long-horizon engineering work — and then, in the same breath, told almost everyone reading that they cannot have it.
Carbon's first users are not chatbot users, not API developers, and not the curious public. They are a subset of vetted cyber defenders, admitted through Google's Fairwind Program. The rest of us get the benchmarks, the pricing, and a wait with no date on it.
I think that is worth writing about anyway: Carbon shows where AI capability is heading, and how the frontier labs now plan to release it. Here is the honest version — what Carbon does, what the numbers really say, why Google put it behind a gate, and what I would do while waiting.
What Google actually announced
Gemini 4 Carbon is the second model in Google's Gemini 4 generation, following Gemini 4 Argon, which arrived on September 30, 2026 — barely four weeks earlier. Two flagship releases in a month would have been unthinkable a couple of years ago. The pace itself is part of the story: Google is clearly in a hurry, and so are its rivals.
Where Argon was a general-purpose flagship, Carbon is a specialist. Google DeepMind's Koray Kavukcuoglu described it as having state-of-the-art capabilities in cyber defense and for long-term, multi-stage tasks. In plain language, Carbon is tuned for the full defensive cycle: find a vulnerability in real software, prove it can actually be exploited, and write a fix that addresses the root cause rather than papering over the symptom.
That specialization matters because security work is where today's AI models hit their limits hardest. Finding a bug in a small demo program is one thing. Tracing a subtle flaw through a massive, messy, real-world codebase, keeping track of everything you have tried, and producing a patch that does not break something else is quite another. Google says Carbon is built for exactly that kind of long, multi-step work, and that its internal teams have watched it run for days at a stretch on complex engineering problems, including hunting subtle bugs in large codebases.
The benchmark triple: find it, prove it, patch it
Google led its announcement with benchmark results, and they are genuinely strong — with one important caveat I will come back to. Carbon's headline scores cover the three disciplines of defensive security work.
First, finding vulnerabilities. On SEC-bench Pro, a benchmark for autonomous vulnerability detection in real-world software, Carbon scores 88.2 percent — a new record. That means it independently identifies complex security flaws, not just textbook examples.
Second, proving a flaw matters. On ExploitGym, which tests whether a discovered vulnerability can actually be exploited, Carbon scores 54.8 percent, again a new record. This step is the one most people skip when they talk about AI security tools, but it is the step that separates a noisy scanner from a useful colleague. A list of a thousand possible flaws is not helpful. A short list of flaws that can really be used against you is.
Third, patching. On CWE-bench, which measures whether a model can turn findings into working fixes, Carbon scores 71.0 percent. Google says it generates secure code at more than twice the success rate of Argon on its internal benchmarks, and that its own Chrome team reported a tenfold improvement in patch quality compared with a model from another AI lab. That rival model was not named, so treat the "10x" figure as a marketing claim, not a measured fact.
On the overall Artificial Analysis Intelligence Index version 5.0, Carbon scores 59 points and ranks first — ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Google says Carbon leads 15 of the 17 benchmark rows in its announcement. The exceptions are worth noting honestly: OpenAI still leads on one software engineering test, and Anthropic leads on one computer-use evaluation. Independent testing also shows Carbon trailing Claude Opus 5.5 on agentic coding, the day-to-day coding assistant work many developers actually use. No model wins everything, and Carbon is no exception.
The coding side: days, not minutes
Carbon is not only a security model. Google says it has also improved general coding and reasoning performance, and the design goal is long-horizon work: agent-based workflows that run for days rather than minutes.
There is one public number for this side of Carbon: on WebDev Arena, a ranking for the quality of web development projects, Carbon jumped 89 Elo points over Argon to land in fourth place with 1,768 points. That is a significant improvement, and it is directly relevant to anyone who builds for the web. It also tells you something about where Carbon sits in the overall race: top-tier, but not a clean sweep.
Why you cannot have it: the Fairwind gate
This is the most interesting — and most debated — part of the launch. Carbon is not in the Gemini app, not in the public Gemini API, and not something you can sign up for today. Google is rolling it out first to a subset of cyber defenders through its Fairwind Program, then to additional trusted partners, and only later to developers, businesses, and consumers. There is no date for any of those later stages.
The Fairwind Program, launched in September 2026 alongside Gemini 3.8 Flash Cyber, covers more than 650 partners according to Google's announcement coverage, including governments, national cyber authorities, and critical infrastructure operators in healthcare, energy, telecom, and finance. It restricts dual-use work to defensive and academic research purposes.
The reason for the gate is dual use, and it is a real reason, not just a press-release excuse. A model that is genuinely good at finding vulnerabilities, writing exploits, and patching code is useful to defenders — and, in the wrong hands, to attackers. Giving the strongest version of that capability to everyone on day one would hand attackers the same tool defenders get, at the same moment. Google's answer is to give defenders a head start.
There is a second reason: safety testing. Google says Carbon is participating in the US government's voluntary pre-release testing process for new AI models, and that broader access will come after it gathers feedback and iterates on its safeguards. That is a sensible sequence, even if it is frustrating for developers who want to test the model themselves.
What this means for developers and WordPress site owners
If you are not a cybersecurity professional, you probably will not touch Carbon directly for a while — and that is fine, because its effects will still reach you.
First, safer software, faster patches. If defenders can find and fix flaws in days instead of weeks, the apps, plugins, and themes you rely on get patched sooner. The Chrome team's patching claim, if even partly true, points in that direction. For WordPress site owners, this is the most practical upside: the long tail of plugin and theme vulnerabilities is exactly the kind of messy, real-world codebase work Carbon is built for.
Second, better AI coding tools down the line. Carbon's long-horizon engineering ability will eventually filter into developer tools and, from there, into the products those developers build. The WebDev Arena jump hints at what that looks like for web work specifically.
What should you actually do while waiting? Nothing exotic. Keep patching on schedule — an unpatched site is still the easiest target there is. Keep your plugin and theme inventory small, and be sceptical of any third-party site claiming to offer "Gemini 4 Carbon access" — if it is not coming from Google, it is not real. A gated launch is a perfect setup for fake-access scams.
The honest caveats
I want to end the analysis with the caveats, because they are what separate a useful article from a press-release rewrite.
One: the headline numbers are self-reported. Google's own data, not yet independently reproduced. AI companies have repeatedly been overly optimistic in their launch materials, and Carbon's numbers deserve the same scepticism as any other lab's.
Two: Carbon is not a clean sweep. Rivals still lead on individual tests, and independent rankings put Claude Opus 5.5 ahead on agentic coding. If your daily work is general coding assistance, the differences between these flagships are smaller than the headlines suggest.
Three: the pricing is real but theoretical for now. Introductory rates are 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens discounted by 95 percent. After the introductory period, Google says prices will rise to 4 dollars and 20 dollars respectively, and it has not said when that period expires. Pricing only matters once you can actually buy access, and for now that door is open only to Fairwind defenders.
Four: the "10x better patches" claim names no rival, and one test reportedly shows a higher hallucination rate than Argon. Both are reminders to wait for independent testing before rearranging your workflow around a model you cannot touch yet.
Conclusion: a specialist that signals where the race is going
Gemini 4 Carbon is not the model most people will use this year — and that is precisely what makes it interesting. Google chose to debut its strongest model yet not as a chatbot upgrade, but as a tool for the people who defend the internet's infrastructure. The benchmarks are impressive, the safety story is encouraging, and the honest caveats — self-reported numbers, no public date, rivals winning individual tests — keep it grounded.
If Carbon delivers on even half of its security promise once independent testers get access, it will quietly make the software you use every day safer. Watch the Fairwind rollout: when Carbon reaches ordinary developers, that is when its real impact begins. Until then, keep your own sites patched, small, and boringly secure.
FAQ
What is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind's frontier AI model announced on October 28, 2026, built for cybersecurity defense and long-horizon engineering. It finds vulnerabilities in real software, validates whether they can be exploited, and writes root-cause patches.
Is Gemini 4 Carbon available to the public?
Not yet. At launch, Carbon is available only to a subset of vetted cybersecurity defenders through Google's Fairwind Program. Google says broader developer, business, and consumer access will follow after feedback and safeguard improvements, but it has announced no date.
How is Gemini 4 Carbon different from Gemini 4 Argon?
Carbon is the specialist successor to Argon: higher on every security benchmark, with secure-code generation at more than twice Argon's rate on Google's internal tests, and able to handle multi-day tasks. Argon remains the general-purpose model most users can access today.
How much does Gemini 4 Carbon cost?
Introductory API pricing is 2 dollars per million input tokens and 10 dollars per million output tokens, with cached inputs discounted 95 percent. Google says prices will rise to 4 dollars and 20 dollars respectively after the introductory period, though it has not said when that period ends.
Why is Google limiting access to Gemini 4 Carbon?
Because its capabilities are dual-use: the same skills that help defenders find and fix flaws could help attackers exploit them. Google is gating first access to vetted defenders while it tests safeguards.
What should WordPress site owners do while waiting for Carbon?
Keep doing the basics that no AI model replaces: update WordPress, themes, and plugins on schedule, remove plugins you do not use, and be sceptical of any site claiming to sell Carbon access. Carbon may make defenders faster, but an unpatched site is still the easiest target there is.
