Google says Gemini 4 Argon can find and patch critical software flaws

Google announced Gemini 4 Argon, its new frontier AI model, and is rolling it out to a set of trusted cyber defenders through its Fairwind Program. Google says the model can locate critical software vulnerabilities, validate them, and patch them without human help, and it will release a version without cyber guardrails to those defenders and to its own internal teams.

Developers, enterprises, and consumers get Argon later, starting with paid API customers and Google AI Ultra subscribers. It says a phased approach is required to release capabilities at this level safely, that it is still adjusting guardrails with feedback from early testers, and that it is taking part in the U.S. government’s voluntary process for pre-release model access.

Before the broad release, Google says it is strengthening safeguards against cyber and chemical, biological, radiological, and nuclear misuse, and that internal and external red teams, groups that attack a system to find its weaknesses, have tested them. It also says monitors watch Argon’s reasoning and actions and stop execution when needed.

What Argon does

Google raised the output limit, the amount of text the model can write in one response, from 64,000 tokens to 1 million. Google says that headroom lets the model think and write for hundreds of thousands of tokens in a single run.

Wiz is using Argon through its Scan for Good program, which scans critical public infrastructure for high-risk exposures and fixes them at no charge. Google says the model found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, one that earlier frontier models had missed. Google did not name the software or say whether it has been patched.

Argon agents applied memory optimizations across Google’s data centers, freeing more than 300 TiB of memory once rolled out, and replaced 32,000 lines of SIMD code in the libgav1 video decoder with Rust, a language built to avoid memory errors. The Rust decoder runs 2.7 times faster than the earlier Rust port with identical video output. Other agents are rewriting C and C++ code in Rust, including more than 800,000 lines for the Fuchsia Zircon kernel, and those rewrites are still in audits, emulation testing, and review before production.

What the benchmarks show

Google reports 77.9% on DeepSWE v1.1, a test of long software engineering tasks, which it calls a state of the art; 51.3% on Zapier’s AutomationBench, where it ranks first; and 91.7% on LVBench, a long-video test, another state of the art. On CWE-bench v1, which tests fixing security vulnerabilities, Argon scores 68% and ties for first.

The vulnerability discovery results come from internal tests. Google’s spans complex codebases in 20 programming languages. In Wiz’s black-box penetration test, the model probes live web systems without source code, and Argon beat 3.8 Flash Cyber at mapping the attack surface, finding vulnerabilities, and producing proof-of-concept evidence.

What it costs

Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, and cached input tokens cost 95% less than the input price. After the introductory period the rates double to $4 and $20.

Download: The Agentic Software Development Guide