What Is Gemini 4 Pro (Argon)? Features, Price, Benchmarks

Google DeepMind announced Gemini 4 Argon on September 30, 2026. The company presents the model as its new frontier model, designed for "complex, long-running workflows": software engineering, enterprise knowledge work such as law and finance, and cyber defense. The most striking number is the output limit: 1 million tokens instead of the previous generation's 64 thousand.

You will also see this model called "Gemini 4 Pro" online. Let's settle that name first, then look at what the model does, what it costs and who can use it.

This post is a news roundup and compilation; I have not tested the model. Right now it is open only to a group of approved organizations. Every number below comes from Google's announcement and independent write-ups, with sources at the end.

Is there a model called "Gemini 4 Pro"?

The official announcement does not mention a Gemini 4 release named "Pro". In Google's post the model is called Gemini 4 Argon. The "Pro" label is shorthand the community uses, based on naming like Gemini 3.1 Pro in earlier generations. Before the announcement, some sites also matched the "Argon" codename with "Gemini 4 Pro"; Google did not confirm that.

The practical upshot: if you are searching for "Gemini 4 Pro", as far as we know today it is Gemini 4 Argon. There is no separate "Pro" price, model page or model ID. For setup details, see How to Install Gemini 4 Pro Argon? Who Can Use It?.

The news at a glance

Feature	Value
Model name	Gemini 4 Argon
Announcement date	September 30, 2026
Developer	Google DeepMind
Input context	1 million tokens
Maximum output	1 million tokens (64 thousand before)
Intro price	$2 / million input tokens, $10 / million output tokens
Standard price (after intro period)	$4 input, $20 output
Cached input	95% discount off the input price
Access	Cyber defenders in the Fairwind Program
Next stage	Paid API customers and Google AI Ultra subscribers, no date

Google has not said how long the intro period will last.

What is it designed for?

Google highlights three areas.

Software engineering. The announcement includes internal examples: one memory optimization effort freed 300 TiB, with total savings expected to land between 500 TiB and 1 PiB. Projects porting C/C++ code to Rust are also cited, with examples over 800 thousand lines. These are Google's own accounts; there is no independent reproduction yet.

Enterprise knowledge work. Legal drafting, financial research and video analysis are listed. On the independent evaluator Vals' index, the model appears in first place.

Cyber defense. The security firm Wiz reportedly used the model to find a critical flaw in healthcare software that exposed personal data, one that earlier models missed. That is also why Google is starting access in this area.

Benchmark results

The numbers from Google's announcement and the write-ups based on it:

Benchmark	Gemini 4 Argon	Comparison
DeepSWE v1.1 (long software tasks)	77.9%	Claude Opus 5.5 74.2%, GPT-6 Astra 74.1%
Vals Index	68.9%	Claude Opus 5.5 67.0%, GPT-6 Astra 63.1%
AutomationBench	51.3%	First place in the announcement
LVBench (video understanding)	91.7%
CWE-bench v1 (vulnerability repair)	68%	Tied first with GPT-6 Astra
Terminal-Bench 4.0	57.4%	Claude Opus 5.5 66.4%
FrontierSWE v2	55.0%	GPT-6 Astra 65.5%

The most important detail is in the bottom two rows: Argon trails on terminal-based and science-type tasks. Its strengths are long context, legal and financial work, and long software tasks. I cover which model suits which job in Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6 Astra.

One more caveat: according to The Next Web, some Google employees have voiced doubts that the model is as good in real workflows as it is on benchmarks. Since there was no independent reproduction on launch day, it is fair to read these numbers as "values reported by Google and the evaluators."

Relationship to Antigravity

Antigravity is Google's agent-focused development environment, announced on November 18, 2025 alongside Gemini 3. Besides the editor view, it has a manager view where you run multiple agents in parallel.

According to Varun Mohan's announcement post, thousands of Google employees are using Argon internally in Antigravity. So the model was first tested in this environment. There is no announcement that Argon can be selected in the public Antigravity release. I will update this post when there is.

Why isn't it public?

Google states the reason openly: because the model's cyber capability is strong, it is opening access in stages. It says it is strengthening its safeguards and its techniques for catching misuse by monitoring the model's internal activity, and that it has joined the US government's voluntary pre-release access process. Partners in the Fairwind Program get a version with relaxed cyber protections.

For general release Google says "as quickly as possible"; according to The Next Web, a senior executive used the phrase "well before the end of the year". There is no firm date.

Who is it relevant for?

• Teams that process long documents and codebases: 1 million input tokens together with 1 million output tokens could make jobs like rewriting a codebase in one pass possible. But remember that output tokens are multiplied by the $10 / million price.
• Security teams: Directly relevant for organizations that can apply to Fairwind.
• Individual developers: Nothing to do right now. Wait for the paid API and Ultra stage.

For how I choose AI tools as an individual developer, see the vibe coding tools comparison.

Frequently Asked Questions

Are Gemini 4 Pro and Gemini 4 Argon the same model?

Officially there is no name "Gemini 4 Pro". The model Google announced is Gemini 4 Argon, and the community attaches the "Pro" label to it.

Is Gemini 4 Argon free?

No. The intro-period price is $2 per million input tokens and $10 per million output tokens. After the intro period it is $4 and $20. Until general access opens, you cannot use it even by paying.

Can I use Gemini 4 Argon right now?

Only cyber defense organizations approved in the Fairwind Program and Google employees can use it. Details are in the access post.

How large is the context window?

1 million tokens for input and 1 million tokens for output. The previous generation's maximum output was 64 thousand tokens.

Can I trust the benchmark results?

The numbers are values reported by Google and the evaluators. There is no independent reproduction yet; do not decide until you have tried it on your own work.

Sources

• Google, Gemini 4 Argon: our next era of frontier intelligence (September 30, 2026): tasks, benchmark results, pricing, output limit, safety rationale.
• Google DeepMind, Fairwind Program: who can apply and the terms of access.
• DataCamp, Gemini 4 Argon: Benchmarks, Pricing, and Access: comparison tables with rival models and availability notes.
• The Next Web, Gemini 4 Argon: Google's new flagship reaches cyber defenders first: release timing and employee doubts.
• Varun Mohan, announcement post (X): internal use in Antigravity.
• Wikipedia, Google Antigravity: Antigravity's release date and features.