Gemini 4 Argon Explained: What It Is, Price and Access

151

Gemini 4 Argon is Google’s biggest model announcement in months, but almost nobody can use it yet. Google revealed it on September 30, 2026, and it is reaching a small group of vetted cybersecurity teams first. This guide explains what it does, what it costs, who gets access, and how much weight to give the benchmark claims. We cross-checked Google’s announcement against Artificial Analysis and trade reporting, and we label anything we could not confirm.

Contents

1. What is Gemini 4 Argon?
2. Can you use Gemini 4 Argon today?
3. How much does Gemini 4 Argon cost?
4. What does the 1 million token output limit mean?
5. How good are the Gemini 4 Argon benchmarks?
6. How does Gemini 4 Argon compare with Claude and GPT-6?
7. Is Gemini 4 Argon as good in real work as its benchmarks?
8. Why is Gemini 4 Argon launching to cyber defenders first?
9. What should beginners do about Gemini 4 Argon now?
10. Frequently Asked Questions

What is Gemini 4 Argon?

Gemini 4 Argon is Google’s new frontier AI model, announced on September 30, 2026, built for long coding, finance, legal and cybersecurity tasks.

Google describes Argon as a model that keeps reasoning through long, multi-step jobs instead of answering one prompt at a time. Koray Kavukcuoglu, SVP at Google DeepMind and Google’s Chief AI Architect, wrote the announcement. Google says thousands of its own staff already use it for specialized coding, deeper research and writing.

Two examples from Google show the idea. Argon agents replaced 32,000 lines of video decoder code in Rust, and the result runs 2.7 times faster than the earlier Rust port with identical video output. Separately, a team of Argon agents found memory savings in Google’s data centers that Google says will free over 300 TiB once rolled out.

If you want an earlier look at Google’s agent work, our post on Gemini’s computer use model covers a model that operates software on your behalf.

Can you use Gemini 4 Argon today?

Not yet. Google is rolling Gemini 4 Argon out first to trusted cyber defenders, with paid API customers and Google AI Ultra subscribers expected next.

Early access runs through Google’s Fairwind Program. Google also says it is taking part in the US government’s voluntary process for pre-release model access while it widens access step by step. It has not given a date for the wider release.

For a beginner, that means you cannot sign up and try it today. Be cautious about any site or app claiming to offer Argon access right now.

How much does Gemini 4 Argon cost?

Introductory API pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 once the launch period ends.

Item	Price (Google, Sept 30, 2026)
Input, introductory	$2 per 1M tokens
Output, introductory	$10 per 1M tokens
Cached input	95% cheaper than the input price
Input, after introductory period	$4 per 1M tokens
Output, after introductory period	$20 per 1M tokens

A token is a small chunk of text, roughly three-quarters of a word. Price per token is only half the story, though. Artificial Analysis measured $1.99 per task on its Intelligence Index at launch prices. Reporting on its data puts GPT-6 Astra at $3.26 per task, and Argon at about $3.98 per task once standard pricing applies.

What does the 1 million token output limit mean?

An output limit caps how much a model can write in one reply. Argon raises it from 64,000 to 1 million tokens.

Google says the extra room lets the model think through hundreds of thousands of tokens in a single run, which helps with hard problems. In practice, that means long reports, large code changes or full legal drafts in one pass.

The catch is cost. Artificial Analysis found Argon produced 110 million tokens across its full test suite, against a median of 82 million for comparable reasoning models. Longer thinking can mean a bigger bill, so a high output limit is a capability to use carefully, not a reason to expect lower costs.

How good are the Gemini 4 Argon benchmarks?

Google reports leading scores in coding, finance and legal work, but they are Google’s own numbers. Independent testing puts Argon level with rivals.

A benchmark is a standard test used to score models, like an exam. It shows performance under test conditions, not how a model behaves in your daily work.

Benchmark	What it tests	Argon score	Source
DeepSWE v1.1	Long software engineering tasks	77.9%	Google
AutomationBench	End-to-end business tasks	51.3% (ranked first)	Google
LVBench	Long video understanding	91.7%	Google
CWE-bench v1	Fixing security vulnerabilities	68% (tied first)	Google
Vals Index	Finance, coding, legal, tax work	68.9% (Opus 5.5: 67.0%)	Vals AI via Investing.com
Intelligence Index	General ability, 10 tests	53	Artificial Analysis

The independent score is the one to anchor on. The Decoder reports that Argon gained 23 points over Gemini 3.1 Pro Preview, which scored 30, a large jump that still stops short of a clear lead.

How does Gemini 4 Argon compare with Claude and GPT-6?

Artificial Analysis scores Argon at 53, level with GPT-6 Astra but behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56.

Those Claude figures come from Implicator’s summary of the index. The same report puts Argon at 57% on Terminal-Bench 4, a coding test, behind Sonnet 5.5 at 64%, Opus 5.5 at 60% and GPT-6 Astra at 59%.

Argon does well on honesty. On the AA-Omniscience test, its hallucination rate was 15%, against 51% for GPT-6 Astra, according to Yahoo Tech. A hallucination is a confident answer that is wrong, which our explainer on why AI makes up fake citations covers. The trade-off: Argon answered 50% of those questions correctly, 13 points below Astra, because it declines to guess more often.

To compare the assistants you can use today, see our ChatGPT vs Claude vs Gemini guide and our Sonnet 5.5 vs Opus 5.5 comparison.

Is Gemini 4 Argon as good in real work as its benchmarks?

Unclear. Bloomberg reported that some Google employees say it struggles with certain real coding tasks, while Google calls that characterization inaccurate.

Per AI Weekly’s summary of the Bloomberg story, the anonymous employees cited problems on some coding jobs, including front-end design. Other people inside Google told Bloomberg the model has caught up with the leading labs.

“It’s a certainty that we are always gonna be at the frontier,” said Koray Kavukcuoglu, SVP at Google DeepMind, in the same report. That is a statement of confidence, not evidence.

The fair reading: Google’s launch numbers are a claim to test, not a result. Independent reruns of the coding scores, once the paid API opens, will settle the question. We have not been able to read the Bloomberg article directly, so we are relying on summaries of it.

Why is Gemini 4 Argon launching to cyber defenders first?

Google trained Argon to find, validate and patch software vulnerabilities, so it is starting with vetted defenders while it strengthens safeguards against misuse.

Google says trusted defenders and its own teams get a version without cyber guardrails. Security firm Wiz is using Argon in its Scan for Good program, and Google says the model found a critical flaw exposing personal data in healthcare software used by hospitals worldwide, a risk earlier frontier models missed.

Before broad release, Google says it is strengthening safeguards in four areas:

• Misuse, including cyber and chemical, biological, radiological and nuclear attacks
• Prompt injection, meaning hidden instructions that try to hijack an AI, where Google says Argon leads on Gray Swan’s benchmark
• Monitoring the model’s reasoning and actions for misalignment
• Hardening the sandboxes used for risky tests

What should beginners do about Gemini 4 Argon now?

Keep using your current assistant, wait for independent hands-on tests, and judge Argon by cost per finished task rather than price per token.

1. Stay put. Argon is not available to you, so there is nothing to switch to.
2. Watch for independent reruns. Look for outside tests of the coding scores once the paid API opens.
3. Compare cost per task. The gap between $1.99 at launch prices and about $3.98 at standard prices shows how a discount can hide the real bill.
4. Test on your own work. If you work in coding, finance or legal tasks, run your own files through any new model. Benchmarks do not include your data.

Frequently Asked Questions

When will Gemini 4 Argon be available to everyone?

Google has not given a date. It says it will widen access as quickly as it can after gathering feedback from early testers and improving safeguards. The first broad groups are paid API customers and Google AI Ultra subscribers. Google is also taking part in the US government’s voluntary pre-release access process.

Is Gemini 4 Argon better than GPT-6 Astra?

On Artificial Analysis’s index they tie at 53. At launch prices Argon costs less per task ($1.99 against $3.26) and hallucinated less (15% against 51%), but it answered fewer questions correctly (50% against 63%). At standard prices Argon’s cost per task rises to about $3.98, which is above Astra’s.

What is the Fairwind Program?

It is the route Google is using to give trusted cyber defenders early access to Gemini 4 Argon. Google says defenders and its own teams get a version without cyber guardrails so they can use the model’s full vulnerability-finding and patching ability.

What does long-horizon mean for an AI model?

Long-horizon means a job with many connected steps that unfolds over a long stretch, such as migrating a codebase or researching a legal question, rather than a single question and answer. Google positions Argon as strong at sustaining this kind of work without losing track of the goal.