Google Announces Gemini 4 Argon, but No One Can Use It Yet

Google announced its newest frontier model, Gemini 4 Argon, on Wednesday — and almost no one can use it. Access is currently limited to cybersecurity defenders in Google's Fairwind program, and the company did not respond to questions about when the model will reach the public.

The restricted launch reflects two forces that have reshaped the AI industry in the 10 months since Google released Gemini 3. The first is policy: the White House introduced a voluntary review process for new AI models, and Google said in a blog post that it is "actively engaged" with it. The second is security. Frontier models now pose an increasingly potent cybersecurity risk to online companies, governments and consumers — a risk made concrete by the string of AI agent hacks seen this summer.

A new release playbook

Google is following a staged rollout pattern that has become standard across the industry. New frontier models go first to cybersecurity experts, who use them to shore up their own defenses and report weaknesses. Subscribers come next, and free users last. OpenAI used the same sequence for GPT-6 Astra, and Anthropic did the same with Fable 5.

The slow-roll release is meant to give experts time to find and patch issues before the AI is let loose on the internet. In past releases, defenders got a few weeks with a model before its general release. Because this is Google's first model to go through the process, the wait could be longer.

The stakes are high. Gemini 4, or a version of it, is expected to become the model powering Google's AI features across Search, Docs and its other products, following the pattern of previous releases. Whatever ships to Fairwind defenders now is likely to shape what billions of users encounter in those products within months.

What Google says the model can do

Gemini 4 Argon handles coding, multimodal tasks such as understanding images and charts, and reasoning. Google said the model is particularly adept at business-related work: software engineering, cybersecurity, and knowledge work in legal and finance fields. Benchmarking data released by Google, comparing Gemini 4 against OpenAI's Astra and Anthropic's Fable 5.1, appears to back those claims, according to the company.

The model has a 1-million-token output limit. More tokens mean the AI can do more work and handle more complex tasks in a single run.

Google also says it has been dogfooding the model internally to make the company more efficient, including in its quantum computing research and its memory operations. Internal deployment before public release has become a common credibility signal among frontier labs, and Google is putting the model to work in two of its most technically demanding divisions.

A launch shadowed by doubts

The announcement follows months of delay and internal friction. Google was rumored to release Gemini 3.5 Pro earlier this year, but reports say it held off to keep working on the model. Even this week, people inside the company were concerned that Gemini 4 was not up to par on key tasks such as coding, Bloomberg reports.

Those concerns cannot be verified until the model ships and independent testers put it through its paces. The benchmark numbers released so far come from Google itself, and the comparison set — OpenAI's Astra and Anthropic's Fable 5.1 — represents the two labs Google is racing directly against for enterprise and developer mindshare.

Why the wait matters

Ten months ago, Google released Gemini 3 and, in CNET's telling, crushed its competition. Since then, the competitive landscape has changed twice over: rivals shipped major frontier models through the new staged-review process, and regulators added a voluntary but consequential checkpoint for releases. Gemini 4 Argon is Google's first model to pass through both gates.

That means the Fairwind phase is not a formality. The defenders now testing the model will determine whether Google's benchmark claims and its internal use cases survive contact with adversarial testing — and how quickly the company can close a release gap that OpenAI and Anthropic have already crossed with their own latest models.

For now, the timeline rests with Google. The company has not said when Gemini 4 Argon will move beyond Fairwind to subscribers and free users, and it did not immediately respond to questions about the schedule. When it does arrive, expect it to spread fast: if past releases are a guide, the model will surface across Search, Docs and the rest of Google's portfolio, carrying whatever the defenders find — or miss — in the coming weeks.