Gemini 4 Argon Goes to Cyber Defenders First. You Wait.

Google announced Gemini 4 Argon on September 30, and the first people to use its new top model aren't developers or Google AI Ultra subscribers. They're a vetted group of cyber defenders in Google's Fairwind Program, and they get Argon with its cyber guardrails switched off. Everyone else, including people already paying Google, has no date. That ordering says more about how dangerous Google thinks this model is than anything in the launch post.

What Is Gemini 4 Argon, and Who Can Actually Use It?

Argon is Google's new frontier model, built for long, multi-step work in software engineering, finance, legal analysis and cybersecurity. Its headline spec is a 1 million token output limit, up from 64,000 on previous Gemini models. That's enough for the model to write or rewrite an entire codebase in one response.

Right now, the only people outside Google using it are Fairwind members. Google says paid API customers and Google AI Ultra subscribers come next, and gives no timeline beyond "rolling out soon."

What the Fairwind Program Actually Is

Google started Fairwind in early September as a limited-access program for organizations that defend networks and software. It already has more than 650 partners. Those include government agencies and national cyber authorities, critical-infrastructure operators in healthcare, telecom, energy and finance, and security vendors such as CrowdStrike, Palo Alto Networks and Wiz. Google is also taking part in the U.S. government's voluntary pre-release access process for frontier models. In Google's words, these defenders and Google's own teams get Argon "without cyber guardrails so they can leverage its full frontier-level cybersecurity defense capabilities."

Why Is Google Holding Gemini 4 Argon Back?

Google trained Argon to find, validate and patch serious software vulnerabilities on its own, and a model that can do that well can also attack. Google is open about this. The wider release will come with safeguards Fairwind members don't get: refusals for requests that would help a cyber or CBRN attack, monitoring of internal activations to spot misuse, stronger resistance to indirect prompt injection, chain-of-thought monitoring for misalignment, and sandboxed execution. Google says it will keep collecting feedback from early testers "as we iterate on guardrails." In other words, those guardrails aren't ready.

What Google Says Argon Can Do in Security

• Autonomous vulnerability work: it finds, validates and patches critical bugs across more than 20 programming languages.
• Black-box penetration testing: according to SecurityWeek's report on the launch, it can probe live systems without seeing their source code.
• A real-world find: Google says Argon found a critical vulnerability in hospital software used around the world. It hasn't named the product.
• CWE-bench v1: 68% on vulnerability remediation, tied for first with OpenAI's GPT-6 Astra and xAI's Grok 4.7.
• Partner use: Wiz is running Argon in its Scan for Good initiative, and Google says Argon beat Gemini 3.8 Flash Cyber on internal vulnerability-discovery tests and Wiz's penetration-testing benchmarks.

Google's top model went to defenders without guardrails before it went to paying customers with them. That tells you Google considers its offensive potential the biggest risk at launch.

Does the $2/$10 Introductory Price Matter Yet?

The pricing is aggressive on paper. Argon costs $2 per million input tokens and $10 per million output tokens during an introductory period, then rises to $4 and $20. Cached input gets a 95% discount. That matches the $2/$10 price OpenAI set for GPT-6.1 Sol, which launched about 24 hours earlier.

The problem is that hardly anyone can pay it. Google hasn't said how long the introductory period lasts or when it starts for the general public. Until paid API access opens, $2/$10 is a number for competitors and analysts to compare. It isn't something a developer can budget around. If the introductory window ends up overlapping mostly with the gated period, the people who benefit from the low price will be the people who had early access.

The benchmarks are just as hard to act on. Google reports 77.9% on DeepSWE v1.1, ahead of Claude Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%. Argon trails Astra on FrontierSWE v2, 55.0% to 65.5%. All of these are vendor-reported, and almost nobody outside Fairwind can check them.

Google Isn't the Only Lab Doing This

Anthropic already restricted its Mythos models to vetted partners through Project Glasswing, so a security-first, invite-only launch is starting to look like the norm for the most capable coding models. What stands out about Google's approach is that it's open about running two versions: one without cyber guardrails for defenders, and one with them for everyone else. Whether that's responsible depends on how carefully Fairwind's 650-plus members are vetted, and Google hasn't published that process.

Key Takeaways

• Gemini 4 Argon is available only to vetted Fairwind cyber defenders and Google's internal teams, and that version has no cyber guardrails.
• Paid API customers and Google AI Ultra subscribers are next in line, but Google hasn't given a date.
• The $2/$10 per million token introductory price (rising to $4/$20) has an unannounced end date, and almost no one can pay it yet.
• Google's own rollout order shows it sees offensive cyber capability as the main risk in its strongest coding model.

Frequently Asked Questions