In this article
Google released Gemini 4 Argon on 30 September 2026, calling it its most powerful AI model yet, but access is limited to selected security partners for now. The model is aimed at coding, research and cybersecurity work, including finding and patching software vulnerabilities.
What Google announced
TechCrunch reports that Google describes Argon as built for deep reasoning across long, complex workflows. Google says it can autonomously find, validate and patch critical software vulnerabilities, support debugging and codebase migrations, and analyse visual content such as video and charts.
The performance claims
Google says Argon scored significantly higher than OpenAI’s GPT-6 Astra and Anthropic’s Fable and Opus models across AI benchmarks, citing Vals benchmarking data that lists it first on the Vals AI index. These are claims by the model’s maker and have not been independently reproduced in the reporting we reviewed, so treat them cautiously until outside evaluators publish results.
Who can use it
Argon is rolling out to select cyber partners through Google’s Fairwind Program and is not available to the general public. TechCrunch says no pricing has been disclosed. The limited release mirrors the cautious approach seen in other cyber-capable models; compare our report on Google, Anthropic and OpenAI’s new cyber AI models and the Nvidia agent safety platform.
Why it matters
A model that can find and fix vulnerabilities on its own is useful for defenders, and the same capability would help attackers if misused. That dual-use tension is the reason for the partner-only rollout. Consider a Vietnamese security engineer at a mid-sized bank: access would depend on joining a partner programme, not on a subscription. Watch for third-party benchmark results, a pricing announcement, and any wider release date from Google. Founders weighing where to build AI companies can read our company formation guide.
Questions readers ask
What is Gemini 4 Argon?
Google’s newest AI model, which it calls its most powerful yet, aimed at coding, research, writing and cybersecurity.
Can I use it today?
Not generally. It is being offered to select cyber partners through Google’s Fairwind Program.
How does it compare with rival models?
Google claims it scored significantly higher than GPT-6 Astra and Anthropic’s Fable and Opus models, but those are the company’s own claims.
How much does it cost?
No pricing has been disclosed, according to TechCrunch.
What can it do in security?
Google says it can autonomously find, validate and patch critical software vulnerabilities.