Gemini 4 Argon: Who Gets It, When

At a glance: Google announced Gemini 4 Argon on September 30, 2026. It is Google’s new top model, aimed at coding, office work such as legal and finance, and cyber defense. You cannot use it yet. It is going first to a small group of security testers, then to paid API customers (developers who pay to build Gemini into their own software) and Google AI Ultra subscribers. In Google’s own table it beats GPT-6 Astra and Claude Opus 5.5 on 13 of 19 tests, but it loses most of the tests where it has to run code or operate a computer.

Gemini 4 Argon arrived the day after OpenAI launched ChatGPT dots, and a week after Claude Opus 5.5 and GPT-6 Sol. We went through Google’s full announcement and tallied its benchmark table ourselves. Every screenshot here comes from Google’s official post.

What is Gemini 4 Argon?

Argon is Google’s first announced Gemini 4 model and the new top of Google’s Gemini family. Google describes it as built for “long-horizon workflows,” meaning jobs where the model works through many steps over a long stretch instead of answering one question.

Google says thousands of its own staff already use it. The examples are big engineering jobs. Argon agents freed up a huge amount of computer memory (more than 300 TiB) across Google’s data centers. They are also rewriting old C and C++ code into a safer programming language called Rust, including 800,000 lines at the core of an operating system. One rewritten piece of video software ran 2.7 times faster than the earlier Rust version.

Nobody outside a small test group can use it yet

Don’t go looking for Argon in the Gemini app yet. It is not there.

Google’s announcement gives this order:

1. Now: “trusted cyber defenders” in Google’s Fairwind Program, plus a small set of testers.
2. Alongside that: the U.S. government’s voluntary process for reviewing models before release.
3. Next: paid API customers and Google AI Ultra subscribers. Ultra costs $100 to $200 a month.
4. Later: developers, businesses and consumers more widely, “as soon as possible.”

Google gives no date for any step after the first. If you use Gemini free or on the $19.99 Google AI Pro plan, you are further back in the queue.

Where does Argon win, and where does it lose?

Google published one table comparing Argon with OpenAI’s GPT-6 Astra, Anthropic’s Claude Fable 5.1 and Claude Opus 5.5.

Argon has the top score on 13 of the 19 benchmarks, ties GPT-6 Astra on one, and loses five:

Test Argon loses	What it measures	Argon	Winner
FrontierSWE v2	Hard coding tasks	55.0%	GPT-6 Astra, 65.5%
Terminal-bench 4.0	Working in a command line	57.4%	Claude Opus 5.5, 66.4%
PostTrainBench	Machine-learning engineering	45.3%	Claude Opus 5.5, 49.3%
Terminal-Bench Science 0.1	Science tasks in a command line	57.6%	GPT-6 Astra, 68.1%
OSWorld-2.0	Operating a computer	69.2%	GPT-6 Astra, 72.6%

All five losses are hands-on tests where the model writes and runs code or operates a computer. That is the area where we found GPT-6 Astra made its biggest jump, and it still leads there.

Argon wins mostly on reading-heavy work such as legal and finance. Its widest lead is on a legal benchmark run by Harvey, a legal AI company: 19.6%, against 6.7% for the next best, Claude Fable 5.1. All four models score low on that test, so read it as “least bad at hard legal drafting,” not “good at it.” Argon also leads on finance research (65.4%), on a long-input reasoning test (84.2% with 256,000 to 1 million tokens of input, where a token is about three-quarters of a word) and on understanding long videos (91.7%).

Keep in mind that Google chose which tests to publish, and nobody outside the test group can check the numbers yet. Google’s evaluation methodology page explains how each score was run.

A million-token answer changes what one request can do

Most model launches talk about how much a model can read, called its context window. Argon’s headline number is about how much it can write. Google raised the output limit from 64,000 tokens to 1 million in a single response.

The old limit was roughly 48,000 words, the length of a short novel. The new one is roughly 750,000 words. Much of that room is meant for the model’s own step-by-step thinking on a hard problem, not for a longer essay back to you.

What will Gemini 4 Argon cost?

For developers, Google set an introductory API price of $2 per million input tokens and $10 per million output tokens. Text you send repeatedly (cached input) is 95% off. Google calls the price “introductory” and has not said what it will be afterwards.

We priced the same job we used for GPT-6 Sol and Luna: summarizing a 30-page report, about 20,000 tokens in and 1,000 out. Our arithmetic from each company’s list prices:

Model	Input / output per million	One report	1,000 reports
Gemini 4 Argon (intro)	$2 / $10	$0.05	$50
GPT-6 Sol	$2 / $10	$0.05	$50
Claude Opus 5.5	$4 / $20	$0.10	$100
GPT-6 Astra	$10 / $50	$0.25	$250

At launch, Argon costs the same as OpenAI’s mid-priced Sol and a fifth of Astra. These are minimums, since a model’s behind-the-scenes reasoning usually bills as output and hard problems use a lot of it. For people using the Gemini app, Google has only said AI Ultra subscribers get early access. Our Gemini pricing guide covers what each plan includes today.

Security testing is why the rollout is slow

Google trained Argon to find and fix security holes in software. For trusted defenders, Google is releasing it “without cyber guardrails,” so they can use all of that ability. The security company Wiz used it in its free Scan for Good program and found a critical flaw in healthcare software used by hospitals worldwide. Google says earlier models had missed it.

The same skill could help an attacker, which is why the public version waits. Google lists four safeguards it is still strengthening:

• Misuse: refusing help with cyber, chemical, biological, radiological and nuclear attacks, tested by internal and outside red teams.
• Prompt injection: resisting hidden instructions planted in web pages and documents. Google says Argon leads Gray Swan’s test for this.
• Misalignment: watching the model’s reasoning and actions and stopping it if it goes beyond what the user asked.
• Sealed test environments: isolating the model during risky training and testing.

If you want the business side of this idea, zero trust for AI agents explains how companies limit what an agent is allowed to touch.

Should beginners care about Argon yet?

No, not this week. You cannot use it, and when it arrives it goes to Ultra and API customers first. Nothing in the announcement is a reason to switch assistants or upgrade a plan today.

It is worth watching if your work is heavy on documents. The benchmark pattern points to long contracts and financial reports. If you are a Google AI Ultra subscriber, check for Argon in the model picker over the coming weeks and try it on the longest document you work with. For jobs where the AI works in a command line or runs a computer, Google’s own numbers show GPT-6 Astra or Claude Opus 5.5 ahead on most tests.

Whichever model tops the table, you still check its work. A model that drafts a contract clause faster still cannot tell you whether the deal is a good one for your business.

Questions readers ask about Argon

Is Gemini 4 Argon available in the Gemini app?

No. At launch it is limited to cyber defenders in Google’s Fairwind Program and a small set of testers. Google AI Ultra subscribers and paid API customers are next, with no date given.

Is Argon better than ChatGPT and Claude?

On Google’s published tests it leads on 13 of 19, mostly knowledge work, long documents and video. GPT-6 Astra and Claude Opus 5.5 still lead on most tests where the model works in a command line or operates a computer. Independent testing will come once more people can use it.

What is the Fairwind Program?

It is the group of “trusted cyber defenders” that Google gives early access to its most capable security models. Google has not published a list of members.

How much does Argon cost?

For developers, $2 per million input tokens and $10 per million output tokens at an introductory rate. Google has not announced consumer pricing beyond saying Google AI Ultra subscribers ($100 to $200 a month) get it early.

What I read to write this

• Google: Gemini 4 Argon, our next era of frontier intelligence (September 30, 2026; rollout, pricing, benchmarks, safeguards)
• Google DeepMind: Gemini 4 Argon evaluation methodology
• Beginners in AI: Gemini pricing (AI Pro and AI Ultra plan prices)
• Beginners in AI: GPT-6 Sol and Luna (OpenAI and Anthropic API prices used in the cost table)

Related guides

• ChatGPT Dots: What They Do: OpenAI’s launch from the day before, which paying Pro users are starting to get.
• GPT-6 Astra: What Really Changed: the model that beats Argon on three of Google’s hands-on tests.
• Claude Opus 5.5: What Changed: Anthropic’s latest, and Argon’s rival on coding.
• GPT-6 Sol and Luna, Explained: the OpenAI model priced the same as Argon.
• Gemini Pricing: Free, Pro and Ultra: what Ultra costs if you want Argon early.
• Claude vs Gemini: Which AI Wins in 2026?: the everyday comparison, before Argon.
• Every AI Model Worth Knowing in 2026: where Argon fits in the wider lineup.