At 4:05 Tuesday afternoon, Google DeepMind posted that Gemini 4 Argon — the frontier model it unveiled on September 30 to “a set of trusted cyber defenders” through its Fairwind Program — is now globally available, rolling out to paid API customers, Google AI Ultra subscribers, and anyone in the Gemini app on a Pro plan.

That reads like the lifting of a two-week embargo. It is actually the completion of a licensing scheme. Nobody wants to say so, because the scheme is dressed as caution.

The Guardrails Were Never the Constraint

Here is the detail buried under the benchmark charts: Fairwind participants received Argon without its cyber guardrails. Everyone else gets the model with them. That is not a safety protocol with a fast lane. That is a two-tier capability market whose differentiator is written permission to do the frightening thing.

Both camps in the AI-safety fight assume the same premise — that guardrails are a brake on the product, and the only live question is how hard to press. Google’s rollout suggests something less flattering: guardrails are the pricing tier. The scary version is a premium SKU, distributed by invitation, with a thirteen-day head start over the sanitized version that reaches the rest of us.

The tracking sites that parsed Google’s launch materials put the terms plainly — $2 per million input tokens and $10 per million output while the introductory period lasts, then $4 and $20. Fairwind participation isn’t priced. It’s selected.

Who Gets the Head Start, and Over Whom

Ask who is actually exposed here and the answer isn’t “society.” It’s a hospital chain’s security team, a mid-sized insurer’s incident-response vendor, the IT shop at a regional utility — organizations whose vulnerability window is now partly set by whether their defender of choice made Google’s list.

A security engineer at a regional hospital network put it to me in a Slack DM the morning after the announcement: “We don’t get to use the good version. We get to be the reason somebody else’s good version gets tested.” He wasn’t being wry about it.

There’s an unflattering implication for people who reflexively distrust regulators. The alternative to a government gatekeeper is not no gatekeeper. It’s a corporate one, with less disclosure and no appeals process. Argon is participating in the U.S. government’s voluntary pre-release review — and simultaneously, an unguardrailed frontier model went to a private guest list. One of those two processes publishes its reasoning.

Self-Reported Scores, Metered Bills

The capability numbers, on their own terms, are modest and self-certified. Argon posts 77.9% on DeepSWE v1.1 and 68.9% on the Vals Index. Claude Opus 5.5, shipped September 22, posts 74.2% and 67.0%. GPT-6 Astra, from September 3, posts 74.1% and 63.1%. Every one of those figures comes from the company selling you the tokens.

The number that isn’t spin is the output ceiling. Google raised it from 64,000 tokens to 1 million — a fifteenfold increase in how much the model can bill you for in a single response, at $10 per million while the promotion runs. A “long-horizon coding benchmark” is measuring, among other things, how many yards of output a task will consume.

And watch where Argon’s introductory price lands when it expires: $4 in, $20 out. That is exactly what Claude Opus 5.5 costs today. Google isn’t undercutting the frontier so much as running a coupon that terminates at parity. GPT-6 Astra, for what it’s worth, is listed at $10 and $50 per million per VentureBeat — the price war is over, and the meter is on.

The story the trade press will run is a horse race: Argon versus Opus versus Astra, who leads this month. The story worth running is that safety, capability, and price have been fused into a single distribution mechanism, and one vendor operates all three valves.

For two years the argument has been a shouting match between people who want the guardrails tightened and people who want them torn off. Google just demonstrated a third option. Sell both. Invite-only, price on request.

Sources