Atlas AI Resources

← Back to AI News

AI products

Google Launched Gemini 4 Argon. Most People Cannot Use It Yet.

Published October 1, 2026

On September 30, Google DeepMind announced Gemini 4 Argon, a new frontier model aimed at hard coding jobs, professional knowledge work, and defensive cybersecurity. Access starts with a small group of trusted cyber defenders through Google’s Fairwind Program. Broader release to paid API customers and Google AI Ultra subscribers is promised later, after more safety work and a U.S. government voluntary pre-release review.

What happened

In a DeepMind blog post, Chief AI Architect Koray Kavukcuoglu said Argon is built for long, multi-step workflows—tasks that take many turns of planning, tool use, and revision rather than a single short answer. Google raised the model’s output token limit to 1 million, up from 64,000. A token is a small chunk of text the model reads or writes; a higher output ceiling means the model can keep “thinking on paper” through a much longer job in one go.

Google reports that Argon scores 77.9% on DeepSWE v1.1 (long-horizon software engineering), 51.3% on Zapier’s AutomationBench (end-to-end business tasks), 91.7% on LVBench (long video understanding), and ties for first at 68% on CWE-bench v1 (fixing security bugs). Those numbers are vendor-reported. TechCrunch and CNBC noted Google’s claim that Argon leads the Vals Index for economic impact across finance, coding, legal, and tax work, and that on some cyber evaluations it ties OpenAI’s GPT-6 Astra and SpaceXSI’s Grok 4.7.

Internally, Google says thousands of employees already use Argon. Examples include beating a published quantum-algorithm baseline by 40% in minutes, finding data-center memory savings of more than 300 tebibytes once rolled out, and helping migrate large C/C++ codebases to Rust— including more than 800,000 lines for the Fuchsia Zircon kernel, under heavy review before production. Cybersecurity firm Wiz is using Argon in its Scan for Good program; Google says it found a critical personal-data exposure in healthcare software that earlier frontier models missed.

Introductory API pricing, when the model opens more widely, is listed at $2 per million input tokens and $10 per million output tokens, with cached input at a 95% discount. After the introductory period, Google says prices rise to $4 and $20 per million. For trusted defenders and Google’s own teams, Argon will ship without cyber guardrails so they can use its full defensive toolkit. Before a public rollout, Google says it is hardening defenses against misuse, prompt injection (hidden instructions that try to hijack the model), and misalignment (the model stepping outside what the user asked).

Why it matters to a normal person

This is Google’s answer, one day after OpenAI’s DevDay, to the race for models that can run long jobs—not just chat. If you follow AI news, the headline is less “a new chatbot” and more “a model Google thinks is strong enough that it will not put it in everyone’s hands yet.” That phased rollout is a practical signal: the lab is betting Argon’s cyber and agent skills are useful to defenders and risky if they spread too fast.

For everyday users, nothing changes this morning. For developers waiting on Gemini, the pricing and Ultra/API path show where Argon is headed. Putting a vulnerability-finding model with vetted defenders first is a different strategy than dumping full cyber capability into a consumer chat app.

What it does not mean

It does not mean you can open the Gemini app and pick Argon today. Fairwind partners and internal Google use come first; consumer and open developer access have no public date. It also does not mean independent labs have confirmed Google’s benchmark table. Most of the scores and internal productivity stories above come from Google, or from press summarizing Google.

It does not mean Argon is “safe because it is limited.” Limited access is a staging choice, not a guarantee. And the cyber examples—including the Wiz finding—are early demonstrations Google chose to highlight, not a full public audit of how often the model is right or wrong on real systems.

Sources