Gemini 4 Argon Can Patch Bugs. Google's Not Sharing Yet.

Google rolled out Gemini 4 Argon on September 30, pitching it as its first truly "long-horizon" model: one built to sit with a hard problem for a while, not just answer a prompt and move on. The headline trick is that it can go hunt for security vulnerabilities in a codebase, confirm they're real, and patch them, all without a human walking it through each step.

Argon isn't showing up in your Gemini app anytime soon. Google is routing initial access through something it calls the Fairwind Programme, and the guest list is short: vetted cybersecurity defenders and Google's own internal teams. Paid API customers and Google AI Ultra subscribers are promised access eventually, but Google hasn't said when, which in product-launch language usually means "once we're sure this won't blow up in our face."

That caution is the actual story here. A model that can autonomously find and verify software vulnerabilities is, by definition, also a model that can autonomously find software vulnerabilities, full stop, for whoever points it at a target. Google says Argon has guardrails that refuse requests that would help carry out a cyberattack or build a chemical, biological, or nuclear weapon, and that it'll keep testing and tightening those safeguards before the gates open wider.

On the DeepSWE v1.1 software engineering benchmark, Argon scored 77.9%, edging out Claude Opus 5.5 (74.2%) and GPT-6 Astra (74.1%). More striking is the output ceiling: Argon can generate up to 1 million tokens in a single response, a huge jump from the 64,000-token limit on Google's previous models, which matters a lot when the task is "read and rewrite an enormous codebase" rather than "answer a question."

Google's own internal use cases make the scale clearer than any benchmark does: engineering teams are using Argon to migrate C/C++ codebases as large as 800,000 lines over to Rust, and to hunt for wasted capacity in data centers, work Google says has freed up somewhere between 500 terabytes and a full petabyte of storage. That's not a chatbot demo. That's a model being trusted with infrastructure most companies would never hand to a junior engineer, let alone an AI.

When Argon does open up more broadly, Google's introductory pricing is $2 per million input tokens and $10 per million output tokens, roughly doubling to $4 and $20 once the introductory period ends. For now, the bigger number to watch isn't the price, it's how long Google keeps Argon fenced off. A model this good at finding security holes being kept on a short leash is Google quietly admitting that "AI that writes code" and "AI that breaks into things" are closer to the same skill than anyone's been comfortable saying out loud.