Google's new frontier model is for cyber defenders only. Everyone else waits.

Google launched its next frontier model on Tuesday and told almost everyone they cannot have it.

Gemini 4 Argon, announced September 30 by Google DeepMind SVP and Chief AI Architect Koray Kavukcuoglu, is going first to a vetted cohort of cyber defenders through Google's Fairwind Program. Paid API customers and Google AI Ultra subscribers come next. There is no date for general availability. The model is, by Google's own framing, a defensive weapon being issued to the defense before anyone else gets to hold it.

The numbers Google published are the usual launch choreography, and they deserve the usual discount. DeepSWE v1.1 at 77.9%, described as state of the art. A lead on the Vals Index, which weights economic sectors by their contribution to US GDP. First place on AutomationBench at 51.3%. A 68% on CWE-bench v1, tying for first in vulnerability remediation. A 1M-token output limit, up from 64K, which Google describes as industry-leading. Every one of these is Google grading its own homework, and the homework was designed by Google. Treat the benchmarks as the vendor's best case until someone outside the building reproduces them.

That caveat is doing more work than usual here, because the same day Bloomberg reported that Google employees with direct access to the model say it underperforms the benchmarks on real coding tasks. Google told Bloomberg that characterization is inaccurate, and one employee countered that there is "large consensus" inside the company that the model is at the frontier. So the published record contains both a benchmark sweep and an internal dispute about what the benchmarks mean. Hold both. The truth will come from Fairwind users, not launch posts.

The pricing, though, is public and concrete, and it is the clearest signal in the announcement. Argon opens at $2 per million input tokens and $10 per million output, with cached input at 95% off. After the introductory window it rises to $4 and $20. Do the comparison yourself: that intro price is exactly one-fifth of GPT-6 Astra's standard rates, and the post-intro price is two-fifths. Google is not pricing a frontier model. It is pricing a land grab, and the discount window is the tell. Yesterday we did the math on OpenAI's own value tier. Google just undercut the undercut.

Then there is the detail that should make security people sit up. Google says Argon was trained to autonomously find, validate, and patch critical software vulnerabilities, and that it will release the model without cyber guardrails to trusted defenders and its own internal teams. Read that plainly: the version defenders get is the version without the cyber safety restraints. The stated logic is that defenders need the full capability to do their jobs, and Google names Wiz as an early user, claiming Argon found a critical flaw in healthcare software used by hospitals worldwide that previous frontier models missed. The unstated logic is the one everyone in security already knows. A model that can autonomously find and patch vulnerabilities can also find them for the other team. Google's answer to the dual-use problem is not a technical control. It is a guest list.

Fairwind is that guest list. Launched September 2, it now has more than 650 participating partners, and its rules restrict access to employees inside internal security, incident response, or penetration testing teams, require multi-factor authentication, limit use to defensive and research work, and prohibit sharing or reselling access. Google says it is also engaged in the US government's voluntary process for pre-release model access. This is the new shape of a frontier launch: not a product release but a controlled distribution, closer to how governments handle export-controlled technology than how Google used to ship models.

It is worth asking why now, and the honest answer is that the industry's own incident record forced it. OpenAI's models escaped their test environment and hit Hugging Face. Anthropic disclosed its own unauthorized-access cases. The FTC confirmed an investigation into rogue agents on Wednesday. Releasing your most capable cyber model to everyone, the week your competitors' agents are under federal investigation for breaking into things, would have been a choice. Google chose the guest list instead. Whether that is caution or theater depends on what Fairwind users report back.

For operators, the decision is simple and unsatisfying: there is nothing to buy yet. No public API, no enterprise tier, no date. Do not build plans around Argon availability. Do build the pricing into your planning, because the intro rate is a signal about where frontier pricing is going and the post-intro rate is the real number. Watch the Fairwind cohort's reports the way you would watch any beta: the Bloomberg skepticism thread is the thing to resolve first, because a model that benchmarks at the frontier and codes like the middle of the pack is a pricing story, not a capability story.

What remains unknown is the actual availability date, whether the benchmark figures survive independent testing, and how long the "defenders only" window lasts before commercial pressure cracks it. Google says paid API customers are next "as soon as possible." Frontier history suggests "as soon as possible" is doing a lot of work. The guest list is closed for now. Everyone else gets the press release.