A security hole that sits in software used by hospitals around the world went unnoticed by earlier frontier systems until a Google model turned it up. Google is now widening access to that model, Gemini 4 Argon.
The most capable build goes first to a small circle of vetted cyber defenders, who will run it with the cyber guardrails switched off. Google’s internal teams get the same access. The rollout runs through Google’s Fairwind Program, with paid API customers and Google AI Ultra subscribers next in line.
Google claims Argon can locate a critical flaw, prove it is genuine, and patch it without human help. The work happened inside Wiz’s Scan for Good, a program that turns Argon loose on public infrastructure that matters, hunting high-risk exposures and closing them at no cost. Google did not name the vendor or confirm a fix.
The model’s output ceiling climbed from 64,000 tokens to 1 million, enough headroom to reason and write across hundreds of thousands of tokens in one pass. Across the company’s data centers, Argon agents tuned memory use and recovered more than 300 TiB once the work was deployed. The same program swapped 32,000 lines of SIMD code in the libgav1 video decoder for Rust, a rewrite that runs 2.7 times faster with identical output. Other efforts are porting more than 800,000 lines of C and C++ in the Fuchsia Zircon kernel, still under audit.
On benchmarks, Google reports 77.9% on DeepSWE v1.1 and 91.7% on LVBench, both called state of the art, and 68% on CWE-bench v1. Pricing starts at $2 per million input tokens and $10 per million output tokens, then doubles when the introductory period ends.
