Google released Gemini 4 Argon this week, its most powerful AI model yet — and the company is keeping it under lock and key while it works with the U.S. government on "pre-release safety evaluations," a move that should set off alarm bells for anyone who cares about who controls the algorithms that will shape what Americans can see, say, and think online.

The stakes are straightforward. The company that dominates frontier AI doesn't just sell software — it decides which questions get answered and which get filtered. Google is positioning Argon as the "safe" and "trusted" option at exactly the moment Washington is looking for a gatekeeper.

Google claims Argon outperforms top models from OpenAI and Anthropic on select benchmarks, particularly in legal reasoning, finance, and cybersecurity. The Artificial Analysis Intelligence Index places it behind only Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5. Tim Law, IDC's director of research for AI, told CNBC the model shows "advanced reasoning on critical tasks" and opens a "competitive front for enterprise knowledge work."

But Fortune reports a different story from inside Google's own walls. Employees who have used Argon say it performs well on benchmarks but disappoints in real-world applications — struggling with certain coding tasks, especially front-end design, and running so large it may be prohibitively costly to deploy. Two insiders described the model as "benchmaxxing" — tuned for test scores over practical usefulness. Google called that characterization "inaccurate."

CNBC framed the rollout as a straightforward comeback narrative. Fortune buried the internal dissent in a brief. Both outlets missed the real question: why is Google launching this model exclusively through what it calls the "Fairwind Program" to hand-picked "trusted cybersecurity partners" while coordinating with the federal government before the public can touch it?

Tulsee Doshi, Google's Gemini model product lead, told CNBC the phased rollout "gives us more confidence" and puts a model "trained and strong in cyber defense in the hands of defenders as soon as possible." Read that carefully — Google decides who counts as a "defender" and who gets access. Everyone else waits indefinitely. The company has given no timeline for a wider release.

Omdia chief analyst Lian Jye Su noted that security incidents from OpenAI and Anthropic models created an opening for Google to "stake its claims as the trusted model provider for safe and secure AI operations." That's the play: competitors fumble on safety, Google rushes in as the responsible steward, and Washington hands it the keys.

Meanwhile, OpenAI fired three employees this week for sharing information with an AI safety group — hardly the industry self-policing we've been promised — and faces a lawsuit over its models' role in a cyberattack on startup Hugging Face. The industry's track record on transparency is abysmal.

Google says Argon is already used internally to optimize data center memory, freeing hundreds of terabytes without new hardware. Maybe. But as IDC's Law acknowledged, "the final proof will be in enterprise production environments once the model is fully released." That proof remains unseen — and there's no date for when Americans will get to judge for themselves.

The AI arms race isn't a contest to serve you. It's a contest to control you. And Google just made clear it plans to decide who gets access to the most powerful tools — with Washington's blessing — before you even know what you're missing.