Google Ships Gemini 4 Argon to Cyber Defenders First, Without Cyber Guardrails

Google's September roundup of its AI releases confirms the month's most consequential launch: Gemini 4 Argon. Google positions the model for real-world software engineering, enterprise knowledge work, and cyber defense, and it rolled the model out first to vetted defenders through its Fairwind program rather than to the general API population.

The capability claims are specific. Unite.AI reports Argon at 77.9% on DeepSWE v1.1, tied for first at 68% on CWE-bench v1 for vulnerability remediation, and 51.3% on AutomationBench for end-to-end business tasks, with the output token ceiling raised to one million tokens, up from 64,000 in prior Gemini models. Introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 after the introductory window.

The release structure deserves close attention from engineers and policymakers alike. Google says it trained Argon to autonomously find, validate, and patch critical software vulnerabilities, and that it will provide the model without cyber guardrails to trusted defenders and its own internal teams, according to Cyber Kendra's account of the launch. That amounts to an acknowledgment that the same model weights usable for patching vulnerabilities are usable for exploiting them, and that Google's chosen safeguard is controlling who gets access rather than training the model to refuse harmful requests.

All three benchmark figures originate with Google. CWE-bench and AutomationBench are young evaluations, and a tie at 68% on vulnerability remediation says little about how the model performs against unseen codebases under adversarial conditions. The month also brought Gemini 3.8 Flash and Flash Cyber, text-to-speech models for voice agents, and reusable Skills in the Gemini app.