Gemini 4 Argon thumbnail with the model name beside a glowing atom graphic on a dark blue background

Gemini 4 Argon Is Google’s Smartest AI Yet, But You Probably Can’t Use It Today

Google has a new top model, and the numbers are strong. Gemini 4 Argon was announced on September 30, and it now matches or beats OpenAI and Anthropic on most of the tests Google published. The catch: unless you work in cyber defense, you can’t use it yet. Here’s what it does, what it costs, and when you might get it.

Quick facts if you only have ten seconds:

  • Announced: September 30, 2026, as the first Gemini 4 model
  • Who has it now: vetted security teams in Google’s Fairwind Program
  • Next in line: paid API customers and Google AI Ultra subscribers, with no date given
  • API price: $2 / $10 per million input/output tokens at launch, rising to $4 / $20 later
  • Headline spec: a 1 million token output limit, up from 64K

What Is Gemini 4 Argon, Exactly?

Argon is Google’s new frontier model and the successor to the Gemini 3.5 line. It’s named after the chemical element, which hints that more “element” models may follow. Google built it for deep reasoning on long, messy tasks: finance work, coding, legal analysis, and long videos.

The standout spec is output length. Most models stop writing after tens of thousands of tokens. Argon can produce up to 1 million tokens in one response. That matters for jobs like rewriting a large codebase or drafting a full report in a single pass.

Google also says it uses this power in-house. According to gHacks, Google teams used it to move more than 800,000 lines of Fuchsia kernel code to Rust.

How Gemini 4 Argon Stacks Up Against GPT-6 Astra and Claude

Google published 18 benchmark results. Per VentureBeat’s breakdown, Argon wins or ties 13 of them. Here are the ones most people will care about:

Test (what it measures) Gemini 4 Argon Best rival
DeepSWE v1.1 (real-world coding) 77.9% 74.2% (Claude)
AutomationBench (business tasks) 51.3% 42.5% (Claude)
LVBench (long video) 91.7% 87.5% (GPT-6 Astra)
Finance benchmark 65.4% 53.5% (GPT-6 Astra)
CWE-bench v1 (fixing security bugs) 68% (tied first) 68% (tied)
Frontier software tasks 55.0% 65.5% (GPT-6 Astra)
Science benchmark 57.6% 68.1% (GPT-6 Astra)

So it isn’t a clean sweep. OpenAI’s GPT-6 Astra still leads on the hardest science and frontier coding tests. But for everyday work like coding, automation, and documents, Argon is now the model to beat.

Independent testers agree it’s in the top tier. Artificial Analysis scores it 53 on its Intelligence Index, tied with GPT-6 Astra. It also measured a 15% hallucination rate, far lower than the 51% it recorded for Astra.

Why You Can’t Use It Yet

This is the part that will annoy most readers. Google is rolling Argon out in stages, and regular users are near the back of the line.

The first group is Google’s Fairwind Program, which gives vetted cyber defenders early access. Google says Argon can find, check, and patch serious software flaws on its own. Through Fairwind, it reportedly found a critical bug in hospital software that older models missed.

That same skill is why Google is being careful. A model that finds security holes could help attackers too. So Google is also running it through the U.S. government’s voluntary pre-release review before opening the doors. It’s a similar playbook to the one Google used for Gemini 3.8 Flash Cyber, just on a bigger model.

Here’s the rollout order Google has described so far:

  1. Fairwind Program security teams (happening now)
  2. Government pre-release review (in progress)
  3. Paid Gemini API customers and Google AI Ultra subscribers (“as soon as possible”)
  4. Wider developer, business, and consumer access (no timeline)

Google hasn’t given a single date for any step after the first. Treat any “launch date” you see elsewhere as a guess until Google confirms it.

Gemini 4 Argon Pricing: Cheap Now, Pricier Later

For developers, the launch pricing is aggressive. Google is offering an introductory rate that’s half the standard price, with no end date announced.

Price per million tokens Intro rate Standard rate
Input $2 $4
Output $10 $20
Cached input 95% off input price 95% off input price

At the standard rate, Argon matches Claude Opus 5.5 and costs well under GPT-6 Astra’s $10 / $50. Artificial Analysis puts its cost at about $1.99 per task at the intro price, roughly 60% of what Astra costs for the same work.

For regular users, the path will likely run through Google AI Ultra, Google’s top paid plan. Free users are going the other way right now. Google is trimming what the free Gemini tier includes, which we covered in our look at the Gemini free tier changes.

Who Should Actually Care Right Now

Not everyone needs to watch this launch closely. Here’s a quick way to think about it:

  • Developers on the Gemini API: watch for the paid rollout and lock in the intro pricing early.
  • AI Ultra subscribers: you’re first in line among consumers, so check the model picker in the coming weeks.
  • Security teams: Fairwind is the only door open today.
  • Free and AI Plus users: nothing changes for you yet.

If you’re a casual Gemini user, the current models will cover your daily tasks fine. Argon is built for long, heavy jobs, not quick questions.

Frequently Asked Questions

Is Gemini 4 Argon available in the Gemini app?

Not yet. Google says Google AI Ultra subscribers will get it first among consumers, but it hasn’t given a date.

How much does Gemini 4 Argon cost?

On the API, it’s $2 per million input tokens and $10 per million output tokens at launch. Standard pricing later doubles that to $4 and $20.

Is Gemini 4 Argon better than GPT-6 Astra?

On most published tests, yes, including coding, automation, finance, and video. GPT-6 Astra still leads on frontier coding and science benchmarks.

What is the Fairwind Program?

It’s Google’s early-access program for vetted cyber defenders. Members can use Argon now to find and fix security flaws.

Will free Gemini users get Argon?

Google hasn’t said. Its current plan starts with paid customers, and the free tier is getting smaller, not bigger.

Our Take

Argon looks like the real deal: top-tier scores, far fewer made-up answers, and a fair price. But a model you can’t use is just a press release. If you’re a developer, get ready to test it the day the API opens, because the intro pricing won’t last forever. Everyone else can relax and wait for Google to name a date.


Comments

Leave a Reply

Your email address will not be published. Required fields are marked *