Google introduced Gemini 4 Argon, a flagship AI model designed for programming, corporate tasks, and cybersecurity. The model supports a context window of 1 million tokens, an increase from the previous 64,000 limit. It is currently being tested by security specialists through the Fairwind program and is part of a voluntary US government safety evaluation.
In performance benchmarks, Argon scored 77.9% on DeepSWE v1.1. On FrontierSWE v2, it achieved 55%, trailing GPT-6 Astra at 65.5% and Claude Opus 5.5 at 62.3%. The model scored 65.4% in Vals Finance Agent v2, 51.3% in AutomationBench, and 91.7% in the LVBench video understanding test. In CWE-bench v1, it reached 68% on pass@1.
Google employees are using the model to optimize data centers, freeing over 300 TiB of RAM with potential savings reaching 500 TiB to 1 PiB. Argon is also assisting in migrating code to Rust, including the Zircon kernel with over 800,000 lines. In one instance, it optimized a quantum algorithm by 40% and improved the speed of a video decoder port by 2.7 times.
API pricing will start at $2 per million input tokens and $10 per million output tokens, with cached input at $0.10. After an introductory period, rates will rise to $4 and $20. The model will eventually be available to API users and Google AI Ultra subscribers, though no public launch date was provided.
Security firm Wiz is testing Argon to identify vulnerabilities. The model successfully detected a critical flaw in medical software that other models missed. Google has implemented a monitoring system to halt tasks that deviate from user intent and continues to refine defenses against malicious prompts.
