Gemini 4 Argon: Google Hands Cyber Defenders the Keys First

Google is not releasing its newest frontier model the usual way. On September 30, Google DeepMind announced Gemini 4 Argon, a model built for long-horizon software engineering, enterprise knowledge work, and cybersecurity defense — and instead of opening it to developers, the company is handing it first to a small set of trusted cyber defenders through a program it calls Fairwind.

The rollout choice is the story as much as the model. Gemini 4 Argon carries Google's biggest capability claims in a year — a 1-million-token output limit, state-of-the-art scores across coding and vulnerability-remediation benchmarks — yet public access is deliberately closed. According to Google, trusted defenders and its own internal teams receive an unguarded version of the model, while everyone else waits. The phased launch makes the safety testing part of the announcement itself rather than an afterthought, a shift analysts say marks a genuine change in how frontier releases are handled.

The announcement came in a post on The Keyword bylined by Koray Kavukcuoglu, senior vice president of Google DeepMind and Google's chief AI architect. He described Argon as delivering frontier performance across complex, multi-step workflows spanning real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense.

A million tokens and a chart-topping benchmark run

The headline technical number is output capacity. Gemini 4 Argon can generate up to 1 million output tokens in a single run, up from 64,000 on the previous generation. Google says the model can produce hundreds of thousands of tokens in a single pass on long problems — the kind of long-horizon work, like migrating an entire codebase, where frontier models have historically broken down. Inside Google, Argon has already been put to work on large-scale codebase migrations, including moving C/C++ code to Rust, according to reporting on the launch.

On benchmarks, Google's chart puts Argon ahead of OpenAI's GPT-6 Astra, Anthropic's Claude Opus 5.5, and Claude Fable 5.1 on many tests — with the caveat that all the figures come from Google. The company reported 77.9% on DeepSWE v1.1, a test of long-horizon software engineering tasks, ahead of Claude Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%. On CWE-bench v1, which evaluates vulnerability remediation against real software flaws, Argon tied for first at 68%. It also took the top spot on Text Arena at 1,525 points, 20 clear of Opus 4.6 High, and ranked first on the Vals Index and AutomationBench at 51.3, with 91.7 on LVBench.

Pricing is aggressive at launch: $2 per million input tokens and $10 per million output tokens during an introductory period, with cached inputs priced 95% off. Google notes the rates double once the intro period expires. Independent analysis drawing on Artificial Analysis data put Argon tied with GPT-6 Astra on the Intelligence Index at roughly 60% of Astra's cost per task — a price-performance argument that could matter as much as the benchmark trophies.

Defenders get the unguarded version

The Fairwind Program is the most unusual part of the release. Trusted cyber defenders receive Gemini 4 Argon without cyber guardrails, the same build Google's internal teams use. Google trained the model for cybersecurity defense specifically: it can autonomously find, validate, and patch critical software vulnerabilities, and Google says it shows impressive leaps over the previous Gemini 3.8 Flash Cyber in discovering attack surfaces and generating proof-of-concept code to validate findings.

The capability already has a concrete result. Wiz ran Argon through its "Scan for Good" program and says the model found a previously unknown critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide — a flaw earlier models missed. Google did not name the affected software.

The launch also folded in process news. Argon went through a voluntary pre-release process with the US government before the announcement, and Google said it is strengthening safeguards against misalignment, misuse by bad actors, and indirect prompt injections before wider access. On that last front, Google's own evaluation puts Argon at the top of Gray Swan's indirect-prompt-injection benchmark.

The skepticism: benchmarks versus real work

Not everyone is convinced the scoreboard tells the whole story. Bloomberg reported that employees with direct access say Gemini 4 does less well when they put it to work on real coding tasks than the benchmarks suggest — a characterization Google disputes. The gap between benchmark performance and internal reception has become a recurring theme in frontier launches, and Argon is no exception.

Some analysts argue the more durable number is not a benchmark at all. Drawing on Artificial Analysis data, coverage of the launch noted a 15% hallucination rate for Argon, the lowest among top-tier models — a differentiator Google itself downplayed in the announcement. If it holds, that would be the kind of reliability metric that matters for the agentic workflows Google is pitching, where a model working unsupervised for hours cannot afford to drift.

Other early independent signals are mixed. Andon Labs added Argon to its Vending-Bench 2 evaluation the same day as launch, where it placed third on simulated year-end cash management — a simulation of long-horizon agent behavior, not a live deployment, but exactly the class of test Argon is meant to excel at.

What the launch strategy says about the industry

Step back and the release pattern is the clearest signal. Frontier-model launches used to arrive with a blog post and an API key. Now the access limits, the named testing partners, the government pre-release process, and the staged guardrail work are part of the announcement itself. One industry observer framed it as a welcome correction after cycles when deployment moved faster than testing.

That shift is visible across the week. Alongside Argon, the same news cycle carried a Senate fight over data-center energy costs, an FTC probe of frontier labs, a White House executive order on "super intelligence," and a lawsuit alleging AI agents hacked third-party sites during safety evaluations. The governance machinery is spinning up at the same speed as the capability curve — and Google's decision to put defenders first, rather than developers, reads as a bet that the next frontier battleground is not who can build the smartest agent, but who can be trusted to aim it.

Wider availability will follow in stages: paid API customers and Google AI Ultra subscribers come next. For now, the most powerful version of Gemini 4 Argon sits in the hands of the people paid to break into software — a deliberate choice, and a telling one, about where Google thinks the risk and the value both live.

Sources: reporting on the September 30 launch by BetaNews, Unite.AI, Real Hacker News, ExplainX, and Inside Telecom; weekly analysis via Another Daily AI Newsletter drawing on AI Weekly, Artificial Analysis, and Bloomberg.