For two weeks, Google’s most powerful AI model sat behind a locked door. Only a handpicked group of cybersecurity defenders — members of something Google calls the Fairwind Program — were allowed to touch it. On Tuesday, the door opened. Gemini 4 Argon is now available to developers, businesses, and anyone willing to pay for Google’s higher-tier AI subscriptions. Here is what that actually means.
Argon is what Google calls a “frontier reasoning model” — the company’s latest attempt to build an AI that doesn’t just answer questions but does sustained, multi-step work. Think less chatbot, more tireless junior colleague. Google built it for what it calls “deep work”: refactoring a massive codebase, grinding through legal documents, hunting for security holes — tasks where the model has to hold context across dozens of steps instead of firing off one clever paragraph.
This is the part that matters most, because the answer is: not everyone. Developers and enterprises get it through the Gemini API and AI Studio starting today. Google AI Pro subscribers — that’s $20 a month — and Ultra subscribers get it in the Gemini app and AI Mode in Search. Ultra alone, the top tier, unlocks it in Google Antigravity, the company’s agentic coding environment. Free users get nothing. Google hasn’t said if or when Argon reaches the free tier.
Google’s claims are specific enough to be interesting. Inside the company, engineers used Argon to refactor six million lines of legacy Chrome code — work measured in weeks instead of months — and say it eliminated an entire category of browser crashes. On YouTube, Argon analyzed backend server data, spotted redundant emoji-rendering code, and cut the frontend comment server’s load in half. It also debugged the flight physics simulator for Wing, Alphabet’s drone delivery company.
On benchmarks, Argon ranks first on the Vals Index, which tests models on real economic work — finance, coding, legal, tax — scoring 68.90%, ahead of Claude Sonnet 5.5 at 67.04%. One rival benchmark, Artificial Analysis, tells a less flattering story: there Argon ties OpenAI’s GPT-6 Astra and trails two Anthropic models. The debate over what AI benchmarks actually prove — and who gets the credit — is older than Argon. Benchmarks, as ever, depend on who’s holding the ruler.
As Google DeepMind’s Koray Kavukcuoglu put it in the launch post: “Argon excels at deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science.”
The API launches at an introductory $2 per million input tokens and $10 per million output tokens — roughly a fifth of what OpenAI and Anthropic charge for their flagship models. Cached inputs are 95% off. The catch: the discount expires November 30, when prices double to $4 and $20. Even then, Argon undercuts its rivals on paper — though cheaper per token doesn’t always mean cheaper per task, since models burn through different numbers of tokens to get the same job done.
Three things make this matter. First, Google spent months trailing its rivals in the frontier-model race — this is its answer, and the aggressive pricing says it wants market share, not just prestige. Second, the battleground has shifted: the fight is no longer about who writes the best poem, but who can do sustained agentic work — code, research, analysis — with the least hand-holding. Google says more than 100,000 of its own employees already use Argon each month, with agentic code submissions up 35% and rollbacks down 30%. That’s Google’s own stat — treat it as a claim, not a fact. Third, access is the story: the best AI is increasingly gated behind monthly subscriptions. The free tier gets the demo; the frontier costs money.
The fine print: Argon’s safety testing isn’t over. The Fairwind Program’s limited rollout was explicitly about gathering feedback and testing safeguards, including a voluntary U.S. government pre-release review — part of a broader shift toward third-party AI safety testing. And the Gemini app quietly gained a “tasks” mode alongside Argon, letting it keep working “even when you’re not at your computer,” connected to more than 40 apps. Alphabet shares ticked up 0.4% to $354.10 on the news.
The AI race used to be about who had the smartest chatbot. Argon suggests the next phase is about who has the hardest-working one — and who gets to use it.
3 Comments