Google's strongest cyber model ships by invitation only

Three decisions about who may use a frontier model landed inside the same 48 hours, and they do not point the same way.

Google began rolling out Gemini 4 Argon only to cybersecurity defenders vetted through its Fairwind Program, with a version stripped of cyber guardrails going to Fairwind members and internal teams, a one million token output limit, and no announced release date for general developers or AI Ultra subscribers (SC Media). The model is finished. The gate is the product decision.

OpenAI spent the same window going the other way, releasing GPT-6 Sol and GPT-6 Luna and cutting their API prices by 50%, to $2 per million input tokens and $10 per million output for Sol and $0.10 and $0.50 for Luna, with Sol at xhigh effort scoring 33.2% on AutomationBench at $0.27 per task against Claude Opus 5's 26.9% at max effort and 11.1 times the cost (OpenAI).

What the narrow gate buys

The strongest argument for Google's choice is on the same day's record under another company's name. A deployment of Anthropic's Mythos cybersecurity model found flaws in which some customer configurations of Epic's MyChart, the system holding more than 320 million US patient records, could let outsiders reach patient data without the intrusion appearing in the software's logs. Epic's founder Judy Faulkner put the resulting pause on most product development at about six weeks (TechCrunch).

The part worth reading twice is that the system could not see the intrusion in its own logs. What found it was a model built for that work and pointed at the right target, which is the entire Fairwind thesis: this class of capability is worth more in a defender's hands than in general release, and the only way to arrange that is to choose the hands.

What it does not buy

The same record shows a gate of that kind failing twice, once in software and once in silicon.

• OpenAI said it disrupted a coordinated distillation campaign that ran through more than 15,000 user accounts from July 1, with request volume spiking on July 24 and 25, and attributed the core cluster of the attempted reasoning theft to China's Moonshot AI, developer of the Kimi model (SC Media).
• Federal prosecutors charged Yiu Kong Lui, 38, of San Gabriel, California with export control violations, smuggling and money laundering over roughly $300 million of Nvidia servers routed through Malaysia and Singapore into China from 2023 until August, with his company paid more than $180 million by Malaysian shipping companies to broker the deals (Business Insider).

A vetting program is a signup surface. OpenAI's was more than 15,000 accounts wide before it closed, and that figure is the point: gates like these are found to have failed in retrospect, by the company that owns them, on its own timetable.

Who sets the gate

Not the agreement signed four days before any of this. The White House Accord on Super Intelligence, signed September 29 by the leaders of Google, Anthropic, Meta, OpenAI, Nvidia and xAI and called "morally binding" by President Trump, contains no penalty for companies that skip its commitments and no requirement to use auditors the companies did not select themselves (Tech Times). It says nothing at all about who may hold a model with its cyber guardrails removed.

The ledger holds one case of an authority outside the labs setting that gate, and it was imposed rather than agreed: OpenAI's GPT-5.6 reached a broad public launch in July 2026 only after the Trump administration lifted restrictions it had placed on the model following a cybersecurity-driven delay (Axios).

So the state of play on October 2. One lab decided on its own to keep its cyber model inside a vetted program. Another halved the price of its flagship on the open API. The only precedent for anyone else making that call is an executive hold, applied and lifted at will. And the document all of them signed asks nothing of either choice.

Capability is not the variable being managed now. Access is, and the companies are the only parties with a hand on it.