Google has introduced Gemini 4 Argon, a flagship AI model designed for complex multi-step tasks in software development, data analysis, finance, law, and cybersecurity. The model features an output limit of 1 million tokens, a significant increase from the 64,000 tokens supported by previous versions.
The company is already utilizing Argon internally to optimize quantum computing algorithms and analyze data center telemetry. These efforts have enabled memory optimization, with potential savings estimated between 300 tebibytes and 1 pebibyte. Additionally, the model is being used to migrate codebases to Rust, including the 800,000-line Fuchsia OS kernel.
In internal benchmarks, Gemini 4 Argon outperformed competitors like GPT-6 Astra and Claude Fable 5.1. It scored 77.9% in DeepSWE v1.1 and 51.3% in AutomationBench. In the CWE-bench v1 cybersecurity test, the model achieved a score of 68%, tying for first place with GPT-6 Astra and Grok 4.7.
Public access to the model is currently restricted. Google has launched the Fairwind testing program for cybersecurity experts to evaluate the model's safety and resistance to malicious instructions. A broader release for developers and corporate clients is planned following this testing phase.
Pricing for the model will start at $2 per 1 million input tokens and $10 per 1 million output tokens, with rates doubling after an introductory period. Cached input tokens will be priced at a 95% discount compared to standard rates.
