Gemini 4 Argon: Google's AI that almost no one can use
Gemini 4 Argon is Google's most advanced model, but only cybersecurity defenders have access. Here's what it does, what it costs, and when it reaches you.
NinaKalkazu.org
You open the Gemini app, look for the new model everyone is talking about, and can't find it. Your phone isn't broken. On September 30, 2026, Google announced Gemini 4 Argon, its most advanced artificial intelligence model, and in the same announcement warned that almost no one will be able to use it for now.
It seems contradictory to launch a product and lock the door. But there's logic behind it, and it says a lot about this moment in AI.
This article explains what Gemini 4 does, why it's restricted, how to read the test table Google released, how much it will cost, and in what order it reaches regular people.
Quick summary: the answer in 30 seconds. Gemini 4 Argon is Google's new flagship model, built for programming, specialized office work, and defense against digital attacks. For now, only partners in the Fairwind security program have access. Next come developers who pay for direct access to the model and subscribers to Gemini's most expensive plan, with no date set. In Google's table, it leads or ties in 13 of 18 tests and loses in 5.
Index
- What Gemini 4 Argon is
- What Gemini 4 Argon does differently
- Why almost no one can use it
- Gemini 4 vs GPT-6 and Claude: how to read the 18 tests
- How much Gemini 4 Argon costs
- When Gemini 4 reaches you
- Google isn't alone: the queue of locked models
- What this changes for you
- Frequently asked questions
What Gemini 4 Argon is
Gemini 4 Argon is Google's new flagship model, announced on September 30, 2026. It was built for long, difficult tasks in three areas: programming, specialized office work (such as legal and finance), and defense against digital attacks. For now, only a group of security partners can use it.
A model, here, is the "brain" behind the app. Think of a car: the Gemini you open on your phone is the body, and the model is the engine. Google swaps the engine from time to time, and Argon is the new engine. For you, this matters because the model defines whether the answer comes out right, shallow, or made up.
The announcement came on Google's official blog, signed by Koray Kavukcuoglu, the new head of Google DeepMind, the company's AI division. The text doesn't explain where the name Argon comes from.
Why it took so long
Google had been without a new flagship model since Gemini 3, from November 18, 2025. That's more than ten months, an eternity in this market. In that interval, OpenAI, owner of ChatGPT, launched the GPT-6 family. Anthropic, owner of Claude, launched new Mythos, Fable, and Opus models.
There was a promise along the way. In a presentation to investors, the company wrote that it expected Gemini 3.5 Pro in June. It didn't come out. According to The Verge, Kavukcuoglu said Google "took a step back" to prioritize the Flash models, which are faster and less powerful.
There was also a change of leadership in August. In a public message from August 5, Google announced that Demis Hassabis left DeepMind's daily operations to become chairman of the division's board and chief scientist at Alphabet, Google's parent company. Kavukcuoglu took over the Gemini models. Argon is the first major launch from this new management.
In one sentence: Gemini 4 Argon is Google's new engine, arrived after ten months without a flagship model, and for now runs only in the garage of a few partners.
What Gemini 4 Argon does differently
The difference is in stamina. Argon was built to handle long tasks with many steps without losing the thread: rewriting an entire system, analyzing a legal case, finding and fixing a security flaw. Google calls this "long-horizon" work and says the model has already changed its engineers' routines.
There are three fronts, according to the announcement:
- Programming: from fixing everyday bugs to migrating entire systems from one language to another.
- Specialized office work: financial research, legal writing, and reading charts, documents, and long videos.
- Cyber defense: finding a flaw in a program, confirming it exists, and writing the fix, all on its own.
One million tokens of output
The most striking number is the output limit. Argon can write up to 1 million tokens in a single response. In previous Gemini models, the ceiling was 64 thousand.
A token (piece of a word) is the unit AI uses to read and write. It works like syllables in dictation: the model doesn't see the entire sentence at once, it advances bit by bit. The word "intelligence," for example, might become three or four tokens. This matters because both the limit and the price are counted in tokens.
In practice, the old ceiling was equivalent to a notebook, and the new one is equivalent to a bookshelf. With 64 thousand tokens, the model had to stop in the middle of a large job and wait for you to ask for the continuation. With 1 million, it can reason through and deliver a long project all at once.
What it has already done inside Google
Google uses Argon internally and cited results in the announcement. Read with a discount: these are examples chosen by the company itself, without outside audit.
- Server memory: multiple copies of Argon, working autonomously, analyzed data from Google's server farms and found adjustments that free up more than 300 TiB of memory (TiB is a computer memory unit). That's equivalent to the memory of about 19,000 16 GB laptops.
- Language swap: the model is rewriting code from C and C++ to Rust, a more secure language. In the core of Fuchsia alone, a Google operating system, that's more than 800,000 lines of code.
- Faster video: in libgav1, Google's video decoding program, Argon changed 32,000 lines and made the Rust version 2.7 times faster, with the same image output.
Google itself warns that these large rewrites still go through audit and testing before going live. Argon proposes, and people verify.
Why almost no one can use it
Because the same ability that fixes a security flaw can be used to exploit it. Google decided to release Argon in phases: first to those who defend systems, then to the rest. In that interval, the company says it will strengthen the model's safeguards. It also delivers Argon to the U.S. government for testing before launch, in a voluntary agreement.
What the Fairwind Program is
Fairwind is Google's restricted access program for digital defense tools. It was launched on September 2, 2026, four weeks before Argon, and had already gathered more than 650 partners worldwide.
According to the program page, three types of organizations are admitted:
- governments and national cybersecurity authorities;
- critical infrastructure operators, such as hospitals, telecommunications, energy, and banks;
- major technology platforms whose software is used by millions of people.
There are rules to stay in. Only internal security teams, incident response teams, and penetration testing teams (simulated attacks that the organization itself conducts to find vulnerabilities) can use Argon. The organization must register who accesses it and protect accounts with two-factor verification.
One detail: not all 650 partners receive the model. The page mentions "a subset" of them with exclusive access to Argon.
The logic of giving the tool first to defenders
Imagine a locksmith who invents a pick capable of opening any lock in the city. If he puts it for sale in the window, thieves buy it the same day. If he delivers it first to lock manufacturers, they have time to reinforce the doors. This is what Google calls an "adaptation window."
The announcement gives a real example. Wiz, a security company using Argon in a free public infrastructure protection project, found a critical flaw in software used by hospitals in several countries. The flaw exposed sensitive personal data. According to Google, previous flagship models had not noticed the problem.
There's a point worth noting. For these partners and internal teams, Google will release Argon without cybersecurity safeguards, precisely so they can use its full defensive capability. The version that eventually reaches the public will have these safeguards.
The four safeguards Google wants to strengthen first
The announcement lists four work fronts before general release:
- Misuse. The model should refuse requests related to cyberattacks and chemical, biological, radiological, and nuclear weapons, without blocking legitimate research.
- Prompt injection. This is when someone hides a malicious command inside a website or document that the AI will read. It works like a fake note from the boss left on the intern's desk. Google says Argon is its most resistant model to this trick.
- Misalignment. This is when the AI, to complete the task, does something the user didn't ask for or want. Think of the employee who deletes files to "organize" the folder. Google will monitor Argon's reasoning and actions and interrupt execution when necessary.
- Test environments. The isolated "rooms" where the model is trained and evaluated will be sealed before high-risk tests.
In one sentence: Argon is locked because it's too good at finding security flaws, and Google wants defenders to have this ability before attackers do.
About caution, Google speaks with moderation. About tests, not so much.
Gemini 4 vs GPT-6 and Claude: how to read the 18 tests
In the table Google released, Argon ranks first alone in 12 of 18 tests and ties in 1. GPT-6 Astra, from OpenAI, wins 3. Claude Opus 5.5, from Anthropic, wins 2. Argon leads in quantity but loses in important programming tests.
These tests are called benchmarks (standardized tests). They're like a college entrance exam for AI: all models answer the same questions and get a score. They serve to compare, but only measure what the test covers.
The count above is from VentureBeat, which broke down the table published in the announcement.
Where Argon wins
The biggest advantages appear in office work and long tasks.
| Test | What it measures | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Harvey Legal Agent | legal research and writing | 19.6% | 5.4% | 3.8% |
| Vals Finance Agent v2 | multi-step financial research | 65.4% | 53.5% | 58.6% |
| AutomationBench | executing business tasks end-to-end | 51.3% | 41.4% | 42.5% |
| LVBench | understanding long videos | 91.7% | 87.5% | 83.7% |
| DeepSWE v1.1 | long-term software engineering | 77.9% | 74.1% | 74.2% |
| CWE-bench v1 (tied with Astra) | fixing security flaws | 68% | 68% | 67% |
Notice the first line. Argon wins by a wide margin but gets less than 20% on the legal test. Being the best doesn't mean being reliable: none of these models replaces a lawyer.
Where it loses
The five losses are concentrated in programming and technical work in the terminal, the black command screen that programmers use.
| Test | What it measures | Gemini 4 Argon | Winner | Difference |
|---|---|---|---|---|
| FrontierSWE v2 | software engineering | 55.0% | GPT-6 Astra, 65.5% | 10.5 points |
| Terminal-Bench Science 0.1 | scientific tasks in terminal | 57.6% | GPT-6 Astra, 68.1% | 10.5 points |
| Terminal-bench 4.0 | acting autonomously in terminal | 57.4% | Claude Opus 5.5, 66.4% | 9 points |
| PostTrainBench | post-training tasks for AI models | 45.3% | Claude Opus 5.5, 49.3% | 4 points |
The fifth loss is also to GPT-6 Astra, in a computer use test that VentureBeat mentions without giving the name or score. The other tests in the table, won by Argon, the article also doesn't detail. Result: the model sold as a programming ace wins DeepSWE but loses FrontierSWE and both terminal tests.
Why to distrust tables made by the model's owner
Whoever builds the table chooses the tests. It's like a team that only publishes statistics from home games: everything there can be true, and the picture still comes out favorable. Google had the honesty to include the losses, which is a good sign, but the selection is still theirs.
There's a second limitation. Since almost no one has access to Argon, no one outside has been able to repeat these tests yet. Until independent developers get their hands on the model, the numbers are a company statement, not a verified fact.
In one sentence: in Google's table, Argon wins the most, especially in office work, but loses in heavy programming and hasn't been tested by outsiders yet.
Performance is half the equation. The other half is price.
How much Gemini 4 Argon costs
Argon will have two prices. At launch, it costs $2 per million input tokens and $10 per million output tokens. After the promotional period, it rises to $4 and $20. Google didn't say how long the promotion lasts (prices checked on 10/02/2026).
This is the API price, the gateway that developers and companies use to connect the model to their own systems. It's not the app subscription. "Input" is what you send the model to read, and "output" is what it writes back. Output costs more because it takes more work.
| Model | Input ($ per million tokens) | Output ($ per million tokens) |
|---|---|---|
| Gemini 4 Argon, launch price | 2 | 10 |
| GPT-6.1 Sol (OpenAI) | 2 | 10 |
| Gemini 4 Argon, regular price | 4 | 20 |
| Claude Opus 5.5 (Anthropic) | 4 | 20 |
| GPT-6 Astra (OpenAI) | 10 | 50 |
Competitor prices come from the official pricing pages of OpenAI and Anthropic, checked on 10/02/2026.
GPT-6.1 Sol is OpenAI's mid-tier model, cheaper than Astra. The table shows Google's strategy. At launch, Argon costs one-fifth of GPT-6 Astra and half of Claude Opus 5.5. After the promotion, it matches Opus, remains well below Astra, and costs double Sol.
There's also a discount for repeated text. When you send the same material multiple times, like a company manual, Google charges 95% less for that part. At launch, that's $0.10 per million tokens.
A good price only helps those who can buy. And here we return to the queue.
When Gemini 4 reaches you
There's no date. Google only says it will release Argon to developers, businesses, and consumers "as quickly as possible" after strengthening protections. What the company disclosed was the queue order, and regular users are at the end of it.
The order, according to the announcement:
- Now: selected partners in the Fairwind Program and Google's internal teams.
- Next: paying API customers and subscribers to Google AI Ultra, Gemini's most expensive plan.
- Later: other developers, businesses, and consumers.
Google didn't say whether Argon will appear in the free app, or with what limits. It also said nothing about specific plans or timelines for Brazil.
Meanwhile, the Gemini you use continues running the Flash line models, which are faster and less powerful. If your question is what you can do without paying, we have a guide about Google's free AI.
Google isn't alone: the queue of locked models
Argon repeats a move Anthropic made before. In 2026, at least two labs held back their strongest model and delivered it first to digital defense teams. The stated reason is the same: the ability to find flaws in programs became good enough to be a risk.
Anthropic debuted Claude Mythos within Project Glasswing, where only partner organizations received the model to find and fix flaws in critical software. The company said it needed to advance safeguards that detect and block the most dangerous results before opening up. We told that story in Claude Mythos under lock and key.
OpenAI went a different route but with the same concern. On August 18, it published a note about the pace of development of its models, saying that new monitoring and security protections began guiding the speed of releases. GPT-6 Astra came out a few weeks later, in September.
The pattern changes the meaning of the word "launch." Until last year, launching meant putting the model in the app. Now, launching can just be announcing, showing a table, and opening a queue. Those following artificial intelligence need to start separating two dates: the announcement and the access.
Some see less virtue in this, and that's our view. Announcing a model no one can test guarantees the headline of leadership without the risk of someone disproving the numbers the following week. Caution can be sincere and convenient at the same time.
What this changes for you
Today, almost nothing changes in your daily life: the Gemini you use is the same as yesterday. What changes is expectation. If you pay or are thinking of paying for AI, Argon becomes a reason not to sign a long contract now. The recommendation depends on your profile.
| If you... | What to do now |
|---|---|
| use free Gemini to study, write, and ask questions | Nothing. Argon has no date for the free plan, and Flash models handle these tasks. |
| subscribe to a paid Gemini plan | Follow Google's announcements. AI Ultra subscribers are at the front of the queue. |
| are a programmer and use AI for code | Wait for independent tests. In Google's own table, Argon loses to competitors in FrontierSWE and both terminal tests. |
| decide technology at a company | Don't migrate because of a table. Save the launch price as a reference and test with your tasks when the API opens. |
| work in cybersecurity at a hospital, bank, energy company, or government | See if your organization fits Fairwind. Access requests are made through a form on the program page. |
One rule applies to all profiles: distrust anyone who says they're already using "Gemini 4" on any website or app. Outside Fairwind and Google itself, no one has access.
Frequently asked questions
Is Gemini 4 Argon free?
It's not yet possible to know. Google only disclosed the API price, aimed at developers and businesses, and announced that Google AI Ultra subscribers will receive the model before others. The company didn't say whether it will enter the free app.
When will Gemini 4 be released to the public?
There's no official date. Google states it intends to release "as quickly as possible" after strengthening the model's protections. The announced order is: security partners, paying API customers and AI Ultra subscribers, and then the rest.
Is Gemini 4 better than GPT-6 and Claude?
It depends on the task. In Google's table, Argon leads or ties in 13 of 18 tests, with a wide margin in legal, financial, and automation work. GPT-6 Astra and Claude Opus 5.5 win in five tests, almost all in programming. No independent testing has been done yet.
What is Google's Fairwind Program?
It's a restricted access program, launched on September 2, 2026, that delivers Google's most advanced digital defense tools to governments, critical infrastructure operators, and major technology platforms. It had more than 650 partners at launch, and a portion of them received Argon.
How much does Gemini 4 Argon cost in the API?
At launch, $2 per million input tokens and $10 per million output tokens. After the promotion, $4 and $20, the same price as Claude Opus 5.5. Google didn't announce the duration of the promotional period (checked on 10/02/2026).
Does Gemini 4 Argon replace the Flash models?
Google didn't say that. Argon is the flagship model, more expensive and aimed at long tasks. Flash models remain the fast, cheap option, and they're what serve most common Gemini use today.
Gemini 4 Argon shows that Google is back in the race for the top, and also that the top now comes with a padlock. Save two questions for any company's next announcement: who can use it, and who outside has tested it. As long as both answers are "no one," the table is a promise.
