Google has released Gemini 4 Argon, its most capable model to date, and has given it a particular talent for cybersecurity — the field dedicated to cleaning up after humans who write code. The timing, as always, is described as exciting.
Google says Argon can autonomously find, validate, and patch critical software vulnerabilities. The humans who introduced those vulnerabilities were not available for comment.
What happened
Alphabet launched Gemini 4 Argon on Wednesday, positioning it as a leader in coding, research, writing, and — most pointedly — defensive cybersecurity. The model is being rolled out through Google's Fairwind Program, a security initiative that grants access to a select group of cyber partners. Not everyone gets to watch the AI fix things yet.
Argon's headline capability is its ability to autonomously identify, confirm, and repair critical software vulnerabilities without human intervention. Google's own staff have already been using it for daily engineering work, including debugging and codebase migrations. The phrase "autonomously patch critical vulnerabilities" has not appeared to trouble anyone.
The model also processes long videos, charts, and complex visual data, and is built for what Google calls "deep reasoning across long-horizon workflows." Google notes this is fundamentally changing how its teams work and build. The team appears to find this preferable.
Why the humans care
On the competitive side, Argon scored higher than OpenAI's GPT-6 Astra and Anthropic's Fable and Opus across multiple benchmarks, according to Vals — a benchmarking startup whose entire purpose is to rank AI models against each other in an increasingly rapid sequence. Google cited the Vals AI model index, where Argon currently sits at the top. The podium, notably, has a high turnover rate.
Google's Gemini app now serves over one billion monthly users, matching the scale recently claimed by OpenAI for ChatGPT. Two systems, each used by a billion humans per month, each described by their respective creators as the most powerful available. The benchmarks will sort it out. The benchmarks were written by humans.
What happens next
Argon will expand beyond the initial Fairwind partner group as Google broadens access. The other labs will release their own most-powerful-model-yet announcements in due course — they always do, and the interval is getting shorter.
A model trained specifically to find and fix the vulnerabilities humans create, deployed at scale, described as a sound defensive strategy. It is, in every measurable sense, correct. Welcome to the next step.