Google Unveils Gemini 4 Argon, Its Most Powerful AI — Why You Can't Use It Yet, and Who Gets It First

"Google just released something huge — can I use it in my Gemini app?" If you asked yourself that this week, you're not alone. On September 30, 2026, Google announced Gemini 4 Argon, a new model that beats the latest versions of ChatGPT and Claude on many tests. The catch: almost nobody can use it yet. Here's why, and what will change once it rolls out.

What is it?

In short, Argon is Google's newest and most capable AI "brain."

Every time you use the Gemini app, a model is doing the thinking behind the scenes. Argon is the new top-tier model in that lineup. The Gemini app reached 1 billion monthly users in August 2026, so Argon could eventually reach a very large audience.

What changed?

First place on most benchmarks

Out of 18 benchmarks Google disclosed, Argon leads outright on 12 and ties for first on one. The comparison set includes OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5.

The biggest gap is on Harvey's legal agent benchmark, which tests drafting legal work. Competitors scored 3.8–5.4%, while Argon scored 19.6%. That's far from perfect, but it signals real progress on specialist knowledge work like contracts.

It doesn't win everything, though. On some coding benchmarks, such as FrontierSWE v2 and Terminal-Bench Science, GPT-6 Astra still comes out ahead.

About 16x more output in a single reply

Previously, long reports would often stop midway with "continue in the next message." With Argon, a full manual-length document can be generated in one go.

Stronger video understanding

Argon scored 91.7% on LVBench, a long-video understanding test. Handing it a meeting recording and asking "just tell me what was decided" becomes more reliable.

How can you use it?

At work: Google says it used Argon internally to migrate codebases of more than 800,000 lines from C/C++ to Rust. One rewritten video decoder ran 2.7x faster.

For security teams: Argon can find, validate, and patch software vulnerabilities on its own. That power is exactly why Google is rolling it out carefully.

What will it cost?

Google says it will expand access "as soon as possible." The order is (1) cyber defenders → (2) paid API customers and Google AI Ultra subscribers → (3) everyone else. No specific dates have been announced.

API pricing for developers is already set. The introductory rate is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. After the introductory period, prices double to $4 and $20.

Summary

Gemini 4 Argon is Google's new flagship model, matching or beating the latest ChatGPT and Claude models on most published tests. Its standout strengths are generating very long documents in one pass and handling specialist work. For now, Google is prioritizing safety and limiting access to a small group of security partners.

It's especially worth watching if you want AI to produce long reports or summarize video at work, or if you're considering Google AI Ultra. For details, see the official Google blog post.