Google's Gemini 4 Argon goes to cyber defenders first

Google is releasing Argon through a deliberately narrow door. The model is going first to the Fairwind Program, which opened Sept. 3 with the smaller Gemini 3.8 Flash Cyber model and has since signed up more than 650 organizations, including CrowdStrike and Palo Alto Networks. Koray Kavukcuoglu, Google's chief AI architect, framed that as a phased approach, and Google says it is still hardening safeguards against misuse for cyberattacks or weapons development. Google is also taking part in the U.S. government's voluntary process for pre-release model access.

The capability story sits in the benchmarks Google published. Argon scored 77.9% on DeepSWE v1.1 for long-horizon software engineering tasks, ahead of Claude Opus 5.5 at 74.2%; AutomationBench showed a wider gap of 51.3% to 42.5%; and VentureBeat counted 12 of 18 benchmarks where Argon led outright. Opus 5.5 still holds Terminal-Bench 4.0, and GPT-6 Astra kept FrontierSWE v2. Google set Argon's output limit at 1 million tokens, up from 64,000 for earlier Gemini models, to match that long-horizon work.

That internal use is already concrete. Google engineers are using Argon agents to move C and C++ code to Rust, including the more than 800,000-line Zircon kernel, and on libgav1 the agents rewrote 32,000 lines as safe Rust, making the decoder run 2.7 times faster than the earlier port with no change to video output. A memory-profiling sweep freed more than 300 tebibytes across Google's data centers. In security, Wiz used Argon in its free Scan for Good program and found a critical flaw in health care software that earlier frontier models had missed.

The open question is timing. Google has not given a date for paying developers and Google AI Ultra subscribers, who are next in line, and Argon's $2 and $10 launch pricing moves to Anthropic's $4 and $20 when it ends. Argon's benchmark lead therefore matters most to the defenders who already have access, while its broader impact hinges on how quickly Google widens the door and on whether the model's safeguards hold up as it does.