Google has unveiled its latest AI model, Gemini 4 Argon, which is currently undergoing limited testing. Internal engineers are already utilizing the model to optimize data center operations, successfully saving 300 TiB of memory. Additionally, Argon agents are actively migrating C/C++ codebases to Rust, covering over 800,000 lines in the Fuchsia OS Zircon kernel and significant portions of the re2 and libgav1 libraries.
The model demonstrates strong performance in benchmarks, achieving a 77.9 percent score on the DeepSWE v1.1 software engineering test, surpassing competitors like GPT-6 Astra, Fable 5.1, and Opus 5.5. It also leads in the Vals Index for economic analysis. While API pricing remains unannounced, Google confirmed that Gemini 4 Argon will support an output limit of 1 million tokens, a substantial increase from the previous 64,000 token limit.
The primary focus for the current phase is cybersecurity. Through the Fairwind Program, select partners like Wiz have gained access to the model. Wiz reported using Argon to identify a critical vulnerability in a global hospital system that other models failed to detect. To address safety concerns, Google has integrated monitoring systems that track the model's chain-of-thought to prevent unauthorized or unsafe actions.
