Gemini 4 Argon: Google's most powerful AI that you can't use yet

Gemini 4 Argon is the model Google introduced on September 30, 2026 and describes as its most powerful yet. For now it only goes to trusted cybersecurity defenders through the Fairwind program, and Google gave no date for opening it to the public.

What Gemini 4 Argon is

According to Google, Argon can find, verify and patch software vulnerabilities autonomously. That means it finds hidden flaws in code, confirms they're real and proposes the patch. You can't use it today: access is for defenders only.

The numbers Google reports

• CWE-bench v1: ties for first place at 68%.
• DeepSWE v1.1: 77.9%, a record result according to Google.
• Output: raises its limit from 64,000 to one million tokens, about 16 times more (our own calculation: 15.6 times).

All figures are results reported by Google, not independent tests.

What isn't known

Google published an introductory developer price, $2 per million input tokens and $10 per million output tokens, but the model isn't open to the public yet. According to TechCrunch, Google claims it beats other frontier models, but the comparison table couldn't be seen, so we don't take it as given. We also don't explain what Fairwind is beyond "trusted cybersecurity defenders".

Why it matters

This is our reading, not a Google line: if an AI fixes flaws on its own, whoever has it first protects the software. The future of cybersecurity wouldn't be watching more, but getting to the flaw first. That's why the defenders-only rollout is the story, as much as the numbers.

How to apply it in your business?

Even if you can't use Argon yet, you can prepare. List the systems you look after, note which have in-house code and which flaws would hurt most if someone else found them first. When Google opens access, you'll know what to ask an AI like this. Meanwhile, check whether your security team qualifies for defender access programs.

For another recent Gemini change, read the full breakdown of Gems becoming Skills.

Frequently asked questions

What is Gemini 4 Argon?

Google's most powerful model by the company's own account, introduced on September 30, 2026, able to find, verify and patch vulnerabilities autonomously.

Can I use Gemini 4 Argon today?

No. For now it only goes to trusted cybersecurity defenders through the Fairwind program, and Google gave no opening date.

How much does Gemini 4 Argon cost?

Google published an introductory developer price of $2 per million input tokens and $10 per million output tokens, but the model isn't open to the public yet.

What results does Google report?

68% on CWE-bench v1 (tied for first) and 77.9% on DeepSWE v1.1, a record according to Google. These are Google's figures, not independent ones.

How many tokens can it write?

Up to one million output tokens; it used to be 64,000.

Conclusion

Gemini 4 Argon shows where AI cybersecurity is heading: getting to the flaw first. For now it's a Google promise with its own numbers and defenders-only access, so it's worth watching closely without treating it as proven.