Google unveils Gemini 4 Argon AI model, restricts initial access

Google has announced Gemini 4 Argon, a new AI model optimized for software engineering and cybersecurity tasks. The company is initially limiting access to a select group of trusted security partners participating in its Fairwind Program to ensure safety before a broader release.

Key details

• •According to The Verge and Ars Technica, Google is currently testing the model internally for tasks such as large-scale codebase migrations and data center memory optimization.
• •Ars Technica reports that the model will support an output limit of 1 million tokens, a significant increase from the 64,000-token limit seen in previous iterations.
• •Google released benchmarking data claiming the model outperforms competing systems like OpenAI's GPT-6 Astra on the DeepSWE v1.1 and Vals Index tests, as noted by The Verge and Ars Technica.
• •The Hacker News reports that Google plans to eventually release a 'guardrail-free' version of the model, though current iterations include monitoring systems to prevent misalignment.
• •As reported by Ars Technica, security partner Wiz has already utilized the model to identify a critical vulnerability in a hospital system that Google claims other frontier models failed to detect.

Background

Google's rollout follows the appointment of Koray Kavukcuoglu as head of DeepMind in August. The announcement comes shortly after OpenAI's recent DevDay, where the company debuted its own new AI agents and models, while simultaneously cautioning against the release of its GPT-6.1 Astra model due to safety concerns.

Why it matters

The phased release approach is intended to allow Google to strengthen safeguards against misuse, including prompt injection attacks. While the model is restricted for now, Google has disclosed API pricing, charging $2 per million input tokens and $10 per million output tokens for early access users.