On Tuesday at 12:01 p.m. ET, Google DeepMind stopped pretending. Gemini 4 Argon — announced September 30, priced at an introductory $2 per million input tokens and $10 per million output — was opened to Google AI Ultra subscribers and paid API customers, ending a stretch in which the most capable model on the public leaderboards was available only through the Fairwind Program: a “limited access” list of 650-plus governments and trusted partners that Google had stood up on September 2 for its cyber-defense tools.

The coverage treated that as a rollout finally reaching its conclusion. The rollout was the product.

A Badge, Not a Bottleneck

Fairwind launched with Gemini 3.8 Flash Cyber paired to Google’s CodeMender agent, with priority for government buyers. Argon was folded into it at the end of September. The model was gated behind organizational identity verification — you didn’t buy your way in, you were admitted, which is a different transaction with a different feeling attached to it.

Here is the part nobody in the cohort was angry about. The credential was the value. A security-ops manager at one of the participating firms, messaging an engineer at a competitor the night before the gate lifted, put it well: “Nobody on the onboarding call asked when it would open up. We were there because we were on the list.”

Run that through the standard libertarian model. If the FCC had published a list of 650 approved firms permitted to use a frontier model, there would be a Federal Register notice, a comment period, and a lawsuit by Friday. A private company published the same list, and it became a lead-generation program with a waiting room full of people who wanted their logo on it. The bottleneck the right spent a decade warning about didn’t need the state to build it. It got built on a Tuesday afternoon by a marketing team.

The Halo Model and the Workhorse

Arena.ai put Argon at No. 1 on its Text Arena leaderboard on October 1 — and eighth on Code Arena: WebDev. The Hacker News thread that followed, 1,330 points and 923 comments, was not a debate about whether the model was secretly already shipping in the Gemini app. It was practitioners noting that the coding harness lags competing tools badly, and that Gemini 3.8 Flash remains the daily driver for most of them.

That tells you what Argon actually is. By Google’s own disclosed results, it leads outright on 12 of 18 benchmarks and trails rivals on Terminal-Bench 4.0 and FrontierSWE v2. Its headline upgrade is a one-million-token output limit, up from 64,000. It is the model you put in the launch post and the sales deck. Flash is the model you put on the invoice.

Which means the Fairwind gate cost Google almost nothing commercially and bought it something a Super Bowl ad cannot: a permanent line in every enterprise procurement document about being a trusted cyber-defense partner. Gating your flagship is cheap when your revenue lives in the mid-tier.

The Introductory Price Is the Whole Story

The number everyone repeated was $2/$10, framed by analysts as a “commodity price floor” because it matches OpenAI’s GPT-6.1 Sol. Two sellers, one price, and a satisfied press release about how competition has driven costs down for everyone.

A floor, though, is a price that holds. This one is scheduled to double to $4/$20 when the introductory period ends. That is not a floor. That is a promotional rate with an expiry date, and the launch coverage wrote the promotional number in the headline because it is the number that sounds like good news.

Two competitors independently arriving at the same figure is what competition looks like when there are only two of them. I’m not alleging anything, and I don’t need to. The buyer’s experience is identical either way. Meanwhile, the maximum-length output — one million tokens at $10 per million — runs about $10 a generation, which is nothing, right up until an autonomous agent loops through that a few thousand times while you sleep. The cheap unit price is the marketing. The aggregate bill is the business.

The Voluntary Process Is the Moat

Google is also participating in the Trump administration’s voluntary process for pre-release model access, according to Technology.org. The reflexive objection on the right is that “voluntary” pre-clearance is regulation by another name — the state gets its hand on the throttle without ever having to pass a statute.

That objection has the incentives backwards. A voluntary process is strictly better for the market leader than no process at all. No process gives you no channel to help write the criteria, no government-adjacent halo to hand your 650 partners, and no paperwork burden that a two-hundred-person startup cannot absorb but a hyperscaler can. Google is not enduring the arrangement. Google is the reason the arrangement is voluntary instead of statutory — because a handshake gets you the moat without the lawsuit.

And that is the part that should bother anyone who thinks the answer to concentrated AI power is deregulation. There is nothing here to repeal. You cannot Federal Register a guest list. You cannot sue a credential. The comfortable theory on the right is that if you take the state out of the room, capability flows to everyone who can pay for it. Google just demonstrated you can build the same bottleneck in private, charge admission to it, and have the customers thank you for the privilege.

Sources