Usually, when Google releases its best AI model, everyone can try it the same day, or nearly.
Not this time.
On 30 September, Google unveiled Gemini 4 Argon, the most powerful model it has ever built. But for now, only a handful of organisations have access to it: experts tasked with defending computer systems.
What Argon can do
Google presents it as a model built for long, complex tasks: programming, carrying out legal or financial research, and defending computer systems. It can produce gigantic responses, up to a million tokens at once, enough to rewrite entire programs.
According to Google, it beats the best models from Anthropic and OpenAI on several tests. As always, these results come from the company itself and await independent verification.
But it is in cybersecurity that it impresses most. Google trained it specifically to find, verify and fix flaws in software on its own. It has already found a critical one in software used by hospitals around the world: a flaw that exposed sensitive personal data, and that previous models had not spotted.
Who has access, and how
Argon is initially distributed through a programme called Fairwind, launched in early September. It is aimed at defenders deemed a priority: governments, healthcare institutions, telecommunications operators, and partners specialising in security.
For them, and for Google's internal teams, the model is delivered without cybersecurity guardrails, so that they can use all of its capabilities. Google does not specify which restrictions are lifted.
Ahead of a wider release, the company says it is strengthening four sets of protections. In particular, it is installing systems that monitor the model's reasoning and actions, and can stop it if necessary. It is also taking part in the US government's voluntary scheme, which allows authorities to test models before their release. Paying customers and subscribers to Google's most high-end offering will come next, with no date announced.
In two months, the three largest AI laboratories have adopted the same logic for their most powerful cybersecurity models. Anthropic reserves its Mythos model for selected organisations. OpenAI gives access to the most advanced uses of Astra only through a programme reserved for defenders. And Google now reserves Argon for its Fairwind programme. An unwritten rule is taking hold: the most dangerous capabilities go first to those who protect.
Why this choice
The logic is simple. As Microsoft's report published the same week shows, attackers are already using AI to move faster. A model capable of finding flaws can be used to fix them, or to exploit them.
By giving it first to defenders, Google hopes they will have time to fix the most serious flaws before models just as powerful fall into the wrong hands. It is a race against the clock, and Google is choosing to give one side a head start.
What we take away
Argon marks a change in tone. A year ago, a new model was released by touting its performance. Today, Google's most powerful model arrives first in the hands of a few carefully vetted defenders, accompanied by a list of protections.
One question remains, which this release raises without answering: who decides that a defender is trustworthy? We devote our reflection this evening to it.