Google Launches Gemini 4 Argon Model for Cybersecurity Defenses with a One Million Token Output Limit

Google DeepMind announced the launch of the Gemini 4 Argon model, an advanced model aimed at software engineering, enterprise knowledge work, and cybersecurity defense. The announcement was published by Koray Kavukcuoglu, Senior Vice President and Chief AI Architect at the company, on Google's official blog.

The company stated that the model achieved 77.9% on the DeepSWE v1.1 benchmark, 51.3% on AutomationBench, 91.7% on LVBench, and tied for first place with 68% on the CWE-bench v1 benchmark for software vulnerabilities. These benchmarks are used to measure the model's ability to write code, run automated tasks, analyze video clips, and discover system vulnerabilities. Each benchmark focuses on a different aspect: the first measures software engineering skills, the second monitors the execution of automated task chains, the third checks comprehension of long video clips, while the fourth assesses the accuracy of identifying known vulnerabilities.

Google raised the model's maximum output limit from 64,000 tokens to one million tokens—a change that allows generating code files and long security reports in a single request instead of splitting them into stages. A token is measured as a unit smaller than a word, so one million tokens is equivalent to very long texts that can reach hundreds of pages.

The model will first be available to verified defenders through the Fairwind program, before moving to a paid API tier and the Google AI Ultra service. On the pricing front, the introductory price is $2 per million input tokens and $10 per million output tokens, which will rise after the promotional period ends to $4 for inputs and $20 for outputs.

What do these terms mean?

Token: The smallest unit of text processed by the model, which may represent part of a word, and is used to calculate cost and request size. Frontier Model: An AI model standing at the highest current levels of performance. Verified Defender: A trusted security entity granted early access to cybersecurity defense tools via the Fairwind program before public availability. Maximum Output Limit: The longest text the model can generate in a single request; the higher it is, the greater its ability to complete a large task at once. Introductory Price: A temporary discounted price offered for a limited time before returning to the base price.