Google has a new flagship AI model, and for once the interesting part is not the model itself. It is the queue to use it. Gemini 4 Argon, announced on September 30, is the first model in Google DeepMind's Gemini 4 generation — and it is not going on general sale. It is going first to a hand-picked group of cybersecurity defenders, while developers, companies and the rest of us wait for later stages that Google has not dated yet.
That roll-out order tells you what this model really is. Argon is built for long, grinding jobs: big software projects, finance and legal research, and hunting down security holes in code before attackers find them. It posts some striking test numbers, it can write far more in one answer than any model Google has shipped before, and its asking price undercuts what you would expect for a top-tier model. Here is what is actually new, what the headline "1 million token output" means in plain terms, and why your name is not on the early list.
What Is Gemini 4 Argon?
Argon is the opening model of the Gemini 4 family, unveiled by Koray Kavukcuoglu, Google DeepMind's senior vice president and Google's chief AI architect. In Google's official announcement, the company describes it as a model "built to sustain deep reasoning across complex, long-horizon workflows." Translated out of launch-day language: it is designed to keep working on one hard problem for a long time instead of answering quick questions.
The target jobs are named clearly — real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defence. This is not a chatbot upgrade for casual chat. Google says thousands of its own engineers already use Argon every day for debugging, codebase migrations and algorithm design, so the model has been earning its keep inside Google before anyone outside saw it.
What a 1-Million-Token Output Actually Changes
Every AI model has a ceiling on how much it can produce in a single run. Google's previous generation topped out at 64,000 tokens. Argon raises that ceiling to 1 million tokens — roughly fifteen times more room before the model has to stop.
A token is about three-quarters of a word, so picture it like this. The old limit could write a long report. The new one could write a short book in one sitting. Google puts it this way: with that kind of headroom, Argon can generate hundreds of thousands of tokens in a single trajectory, thinking through a tough problem in one continuous pass instead of being chopped into pieces that lose the thread.
Three kinds of work open up when the output budget gets that big:
- Whole-codebase changes in one pass. Instead of editing file by file and hoping the pieces still fit, the model can rewrite or migrate a large codebase as a single job. Google says Argon has already converted C and C++ code to Rust in parts of its Fuchsia operating system, including a 32,000-line block of video-decoder code that came out running 2.7 times faster.
- Agent runs that do not stall. Long tasks — inspect, reason, write, test, fix, repeat — can stay in one unbroken chain. That is the difference between an assistant that starts a migration and one that finishes it.
- Long professional documents with the working shown. Full audit reports, redrafted contracts, research memos: output length stops being the reason the answer gets cut short.
One honest caveat. A bigger output budget makes longer jobs possible; it does not automatically make every answer smarter. Whether quality holds steady across a very long run is exactly what early testers are about to find out.
The Benchmark Numbers, Read Carefully
Google's chart puts Argon ahead of its two biggest rivals on the tests that match its pitch. Take the figures seriously, but take them as Google-reported: some scores were measured in-house, and rival numbers were drawn from public leaderboards, so the comparison is not perfectly controlled. Even TechCrunch's launch coverage, which called Argon Google's most powerful model yet, leaned on Google's own claims. There is also a counter-signal worth knowing: Bloomberg reported that some Google staff found Argon's day-to-day coding less impressive than its test scores, a characterisation Google disputes.
| Benchmark | Gemini 4 Argon | What it tests |
|---|---|---|
| DeepSWE v1.1 | 77.9% (Claude Opus 5.5: 74.2%, GPT-6 Astra: 74.1%) | Long, real-world software engineering tasks |
| AutomationBench | 51.3% — first place | End-to-end business and computer automation |
| Vals Index | Top spot | Finance, legal and tax knowledge work |
| CWE-bench v1 | 68% — tied for first | Finding and fixing real software vulnerabilities |
| LVBench | 91.7% | Understanding very long videos |
The fair summary: Argon sits in the top tier on paper, with its clearest lead in exactly the areas Google built it for — long coding work and enterprise knowledge tasks. Whether it is truly number one depends on whose test bench you trust, and independent runs will settle that over the coming weeks.
What It Costs — and Why Free Users Are Not in the Queue
Google has published API pricing even though almost nobody can buy access yet, and the numbers are aggressive for a flagship. Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off. When the introductory period ends, the rates step up to $4 and $20.
| Gemini 4 Argon at a glance | Details |
|---|---|
| Maker | Google DeepMind (first Gemini 4 model) |
| Announced | September 30, 2026 |
| Max output | 1 million tokens (up from 64,000) |
| Built for | Long-horizon coding, finance and legal work, cyber defence |
| Intro pricing | $2 per million input tokens / $10 per million output tokens |
| Standard pricing later | $4 / $20 per million tokens |
| First access | Fairwind Program (trusted cyber defenders) |
| Next in line | Paid API customers, then Google AI Ultra subscribers |
| General API date | Not announced |
Notice who is missing from that list: free users. Argon is a paid, gated product from day one, which fits the direction Google has been moving — in the same week, it tightened what the free tier gets, as we covered in our piece on Google Gemini free users losing Flash and Pro on October 9.
Who Gets It First — and Why Cyber Defenders?
The first outsiders through the door are not app developers or big banks. They are the people who defend networks. Google's Fairwind Program, launched in September, brings together more than 650 organisations — government bodies, critical-infrastructure operators and security firms such as CrowdStrike, Palo Alto Networks, Snowflake and Wiz. Those trusted defenders get Argon first, and Google says they get it without the usual cyber guardrails.
The reason is the model's sharpest skill, which cuts both ways. Argon was trained to dig through code, find critical vulnerabilities, confirm they are real, and patch them on its own — that is what the 68% score on CWE-bench is measuring. Pointed the wrong way, the same talent is an attack tool. So Google is placing it with defenders first, watching how it behaves in real environments, and tightening safeguards before the gates open wider. The early returns match the plan: security firm Wiz says Argon has already picked out a critical flaw in hospital software that earlier AI models missed.
There is a second gate as well. Google says it is going through the US government's voluntary pre-release access process in parallel, so officials get a look before the commercial stages begin. The sequence, in order: Fairwind defenders, paid API customers, Google AI Ultra subscribers, and only then the wider public. Handing a powerful model wider permissions is a theme running through Google's recent launches — our breakdown of the Gemini Full Access permission and what it unlocks shows how much machine access these tools now ask for.
When Does Everyone Else Get It?
Short answer: nobody outside Google knows the date, because Google has not given one. The company says it will gather feedback from early testers, keep improving guardrails, and make Argon available to developers, enterprises and consumers "as soon as possible." Published pricing usually means the paid API is not far behind the defender phase — companies rarely print a price list for something nobody can buy for a year — but that is reading signals, not a promise.
For most readers, the practical takeaways are simple. If you work in security, check whether your organisation qualifies for the Fairwind Program; that is the only front door right now. If you are a developer, expect the paid API next and plan around the introductory pricing while it lasts. And if you are a free Gemini user, Argon changes nothing for you today — it is a look at where Google's paid top end is heading, not a feature coming to your app this month.
Frequently Asked Questions
What is Gemini 4 Argon?
It is Google DeepMind's first Gemini 4-generation model, announced on September 30, 2026. It is built for long-horizon work: big coding projects, enterprise research in fields like finance and legal, and cybersecurity defence.
What does the 1 million token output limit actually mean?
Argon can produce up to 1 million tokens — roughly 750,000 words — in a single run, up from the previous 64,000-token limit. That lets it rewrite whole codebases, run long agent tasks without stopping, and draft full-length professional documents in one pass.
Who can use Gemini 4 Argon right now?
Only trusted cybersecurity defenders in Google's Fairwind Program, alongside Google's own teams and a US government pre-release review. Paid API customers and Google AI Ultra subscribers are next. There is no public sign-up and no announced date for general access.
How much does Gemini 4 Argon cost?
Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached inputs 95% off. After the introductory period, the rates rise to $4 and $20 per million tokens.
Is Gemini 4 Argon really better than GPT-6 Astra and Claude Opus 5.5?
On Google's published chart, yes: 77.9% on DeepSWE v1.1 against GPT-6 Astra's 74.1% and Claude Opus 5.5's 74.2%. All of those figures are vendor-reported, though, so the ranking is worth treating as provisional until independent testers publish their own runs.
0 Comments