A locksmith who invents a better pick has an awkward problem. The pick does not know whose door it is in. It turns just as smoothly for the burglar as for the owner, and no amount of good intention on the locksmith's part changes the mechanics.
That, roughly, is the problem Google was describing on 28 October, when it announced Gemini 4 Carbon as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." The cybersecurity part is the interesting one. Google says Carbon is unusually good at three jobs: finding weaknesses in software, checking whether they can really be exploited, and writing the repair. Those skills arrive as a set. A model that can find a hole and prove it is real is, by the same act, a model that could help someone climb through it.
So Google did something slightly unusual with it. The first people to receive Carbon were a subset of cyber defenders. Paying developers, businesses and the public waited.
Why the defenders went first
The logic is a race, and a head start. A flaw in widely used software exists whether anyone has found it or not. Once a capable tool for finding such flaws is in everyone's hands, the question becomes who finds each one first: the people who will patch it, or the people who will use it. Giving defenders the tool early tilts that race, for a while, in their favour.
Google routes this through its Fairwind Program, which it says gives trusted partners "a critical head start against cyber threats." The priority list tells you how Google thinks about the stakes: governments and national cyber authorities, operators of healthcare, telecommunications, energy and financial networks, and the large technology platforms that much else depends on. Applicants go through background checks, access is limited to internal security teams, and handing it on to anyone else is forbidden.
Google also says it is "actively engaged" in the United States government's voluntary process for testing models before release. What that testing found, the announcement does not say.
What a staged release is
A staged release lets a powerful thing out in rings, the smallest and most trusted first, and watches what happens at each ring before opening the next. Google's own words were "a deliberate approach to rollout," with time to gather feedback and adjust its guardrails, the rules and filters wrapped around the model to stop it doing what it should not.
The second ring opened on 10 November. Google's availability post makes Carbon available to paying developers through the Gemini API and Google AI Studio, to companies through Gemini for business, and, in a rollout still under way, to Google AI Ultra subscribers in the Gemini app and AI Mode in Search. The introductory price matches Gemini 4 Argon's: $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 once the introductory period ends on 30 November.
One detail is easy to miss, and I think it matters most. The staging did not really end. Google's model card says Carbon's specialised dual-use cyber capabilities remain available only to verified defenders through Fairwind, and the pricing applies, in Google's footnote, to the Carbon model intended for public access. The version you can now use and the version defenders use are not quite the same. The gate has moved from who may use the model to what the model is permitted to do.
As for Argon, it is the model Google released two weeks before Carbon, and the model card says Carbon is based on it. Google now presents them side by side. Argon, it says, "remains the ideal choice for the majority of software engineering and knowledge work," while Carbon is meant for the hardest, longest problems, and spends more of its effort exploring and checking before it commits to an answer. A fair rough picture is Argon given more patience and pointed at the difficult cases.
The honest limits
Almost everything we know about Carbon's abilities comes from Google, or from partners Google chose to quote. The benchmark results, the claims about safer code, the reports from firms such as Wiz: these are the maker's account. Google itself notes that its external cyber benchmarks were "evaluated without guardrails," which means they measure the raw engine, not the product people actually use. That does not make the claims false. It does mean nobody outside has yet checked them, and coverage at the time, such as Techzine's, largely repeated Google's figures.
The head start also has a shape worth seeing clearly. From 28 October to 10 November is less than a fortnight. Google reports that in that time its own vulnerability researchers used Carbon to build a proof that a flaw in a popular open-source project could really be exploited, which helped the project's maintainers rate it as critical and issue a patch quickly. That is real, and it is good. But finding a flaw is the fast part. Writing the fix, testing it, shipping it, and waiting for thousands of organisations to install it is slow, human work, and no model can hurry a hospital into updating its systems. Two weeks is enough to find things. It is rarely enough to fix them everywhere.
Then there is the question no one can answer from outside: whether the gate holds. Vetting, logging and strong sign-in rules raise the cost of misuse. They cannot make it impossible.
So does staged release work? Partly, and in a particular way. It cannot stop a capability from spreading. What it can do is decide who practises with it first, and let the early lessons happen where mistakes are cheapest. That is a modest ambition, and possibly the only honest one available.
The pick still does not know whose door it is in. Google has simply tried to make sure the owners heard about it first.
Sources
Gemini 4 Carbon: our next frontier model, Google
Gemini 4 Carbon is now available, Google
Gemini 4 Carbon model card, Google DeepMind
Fairwind Program, Google DeepMind
Google Gemini 4 Carbon is fully committed to cyber defense, Techzine
