The History of AI

Gemini 4 Argon

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models

Lineage
Neurons and networks
Kind
Implementation

Who

Google DeepMind

What happened

Google presents Gemini 4 Argon, its most capable frontier model, and gives it first only to selected cybersecurity defenders, through its Fairwind program. It is the first new generation of Gemini since Gemini 3.

Why it mattered

Google adopts the risk-based staggered rollout for cybersecurity that Anthropic and OpenAI had already followed. With it, all three major American labs release their most capable model to defenders first.

The full story

In 2026 the most capable models found software vulnerabilities better than almost any person. Anthropic had reserved Claude Mythos Preview for Project Glasswing in April, and OpenAI had limited access to GPT-6 Astra in September. Google followed the same path.

On September 30, 2026, Koray Kavukcuoglu, Google's chief AI architect, presented Gemini 4 Argon, designed for software engineering, office work in fields such as law or finance, and cyber defense. It went first to members of the Fairwind program, which Google had opened on September 3 with a smaller model and which already brought together more than 650 cybersecurity organizations. Google also took part in the US government's voluntary process for reviewing models before release.

According to Google, Argon led most of its benchmarks. It scored 77.9 percent on DeepSWE, for coding, ahead of Claude Opus 5.5 and GPT-6 Astra, and tied with Astra on CWE-bench, which measures how well vulnerabilities are fixed. Its responses could run to a million tokens, against 64,000 in the previous generation, to sustain long chains of reasoning.

On safety, the company said it followed its Frontier Safety Framework: the model refuses to help with cyberattacks and with chemical, biological, or nuclear weapons, and it monitors its own reasoning and actions so that it can stop if something goes off course. It would later reach developers, businesses, and paying users, at $2 per million input tokens and $10 per million output tokens during the initial period.

Builds on

  • Claude Mythos Preview and Project Glasswing (April 7, 2026), GPT-6 Astra (September 3, 2026)

    Inherited, with no proven direct link

    Gemini 4 Argon goes first to selected cybersecurity defenders, the access scheme Anthropic introduced with Glasswing and OpenAI followed with GPT-6 Astra

Threads that pass through here

Sources

How to cite

“Gemini 4 Argon.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/gemini-4-argon

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline