Gemini 4 Argon: what Google announced, who can use it and what it will cost
Google's new AI model, Gemini 4 Argon, is only open to a small group of trusted cyber defenders for now. We explain what Google announced, who is next, and how its $2 / $10 price compares.
On 30 September 2026 Google announced Gemini 4 Argon, its newest top AI model. The numbers sound big. Google says Argon can write up to one million tokens in one answer. We did not test that limit. Tokens are the small pieces of text an AI reads and writes. Google also says Argon got top scores on several tests Google picked. And the starting price is $2 per million tokens in and $10 per million tokens out. But there was a catch at first: only a small group of trusted cyber defenders, teams that protect computer systems, could use Argon. Since 13 October it is in the paid Gemini API, see the updates below. We read the announcement, the documents for developers and the price pages of four AI companies. We wanted to answer three questions: what did Google announce, who can use it, and what will it cost?
What did Google announce?
The announcement was written by Koray Kavukcuoglu. He is a senior vice president at Google DeepMind and Google's chief AI architect. Google calls Argon "our next era of frontier intelligence". It aims the model at three kinds of work: writing software, office work such as law and finance, and protecting computers against attacks.
Most of the post is about how Google already uses the model. Google says thousands of its workers use it. It gives two examples. First, the model helped its data centres use memory better, which freed "over 300 TiB of memory". Second, AI helpers reworked a free video program written in the Rust language. Google says that program now runs 2.7 times faster. Outsiders cannot check these results, so we report them as Google's claims.
Before the announcement, stories about a model called Argon went around online. They do not agree on when it first appeared. We could not confirm either date, so we leave that story out.
Who can use it, and when?
Most people cannot use it yet. Google says Argon "is rolling out to a set of trusted cyber defenders through our Fairwind Program". Fairwind is a security programme Google started in September 2026. It is for governments and companies that protect computer networks. It began with an older cyber model, Gemini 3.8 Flash Cyber. Google's own teams also use Argon.
Google calls this "a phased approach": it opens the model up step by step. Google says it takes part in "the U.S. government's voluntary process for pre-release model access". It will first improve its safety rules with early testers. After that, Argon will go to "developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers".
Two things are missing from that sentence: a date and a region. Google does not say when the next step comes. It does not mention Europe or any other region either. The author of this site lives in the Netherlands. People there cannot know yet if Argon will arrive at the same time as in the US.
What will it cost?
Google gives two prices, in US dollars per million tokens:
| Text in | Text out | Saved text in | |
|---|---|---|---|
| Starting price | $2 | $10 | 95% off the text-in price |
| After the starting period | $4 | $20 | not stated |
"Saved text in" is text the model already read before and can reuse at a lower price. The second row comes from a small note under the announcement. You can see it in the screenshot in this section. Google does not say how long the starting period lasts. Update, 17 October: Google's price list now says the starting price runs until 30 November 2026. From 1 December it is $4 in and $20 out per million tokens.
How does that compare? On the same day we looked up the normal prices for developers at Anthropic, OpenAI and xAI. The chart at the top of this article shows the same prices.
| Model | Text in | Text out |
|---|---|---|
| Claude Opus 5.5 | $4 | $20 |
| Claude Sonnet 5.5 | $2 | $10 |
| GPT-6.1 Sol (up to 272K tokens in) | $2 | $10 |
| GPT-6 Astra | $10 | $50 |
| Grok 4.7 | $2 | $6 |
| Gemini 3.8 Flash (until 31 Dec 2026) | $0.75 | $3.75 |
At the starting price, Argon costs the same per token as Claude Sonnet 5.5 and GPT-6.1 Sol. After that, it costs exactly the same as Claude Opus 5.5.
But the price per token is only half the story. A model that thinks longer writes more tokens for the same job. Google's bigger writing limit invites exactly that. We saw this in our own tests. In Codex, GPT-6 Astra used about six times the credits of GPT-6.1 Sol for the same result. Credits are the units a subscription's allowance is counted in. And Grok 4.7 cost twice as much as Grok 4.6, at the same price per token.
Can it really write one million tokens at once?
The clearest technical change is how much the model can write in one answer. Google says it is raising that limit "to an industry-leading 1M tokens, up from the previous 64K tokens". You can compare it with Google's own page for Gemini 3.8 Flash. That page lists a limit of 65,536 tokens. That is the 64K Google means. A bigger limit lets a model write or think much longer in one go. That helps with big jobs, such as moving a large program to new code.
Google also shares test scores. On DeepSWE v1.1, a test of long programming jobs, Argon scores 77.9%. On AutomationBench, Zapier's test of business tasks, it scores 51.3%. On LVBench, a test of understanding long videos, it scores 91.7%. These are Google's numbers on tests Google chose. We could not repeat them.
It is worth reading Google's table to the end. FrontierSWE v2 is another programming test in the same table. There Argon scores 55.0%. That puts it behind GPT-6 Astra (65.5%), Claude Opus 5.5 (62.3%) and Claude Fable 5.1 (56.3%). So a model can win one programming test and lose another.
Why only cyber defenders? Fewer safety limits for some
Computer security is the reason for this unusual start. Google says Argon can "autonomously find, validate, and patch critical software vulnerabilities". In other words: it can find serious security holes in software, check them and fix them on its own. For trusted defenders and its own teams, Google says it will release Argon "without cyber guardrails". That means without the usual safety limits, so they can use everything it can do.
That explains the order. A model that is very good at finding security holes helps defenders. But it could help attackers just as much. Google says it is making its protection against misuse stronger before a wider release. For example, it watches the model's reasoning for actions that go beyond what the user asked.
Google also names an outside user: the security company Wiz. Google says Wiz uses Argon in a free programme to find unprotected systems in public services. On CWE-bench v1, a test of fixing security holes, Google says Argon "ties for first place" with 68%.
Can developers find it yet? Not at first
We checked if developers can find Argon where they would normally look. We did this on 4 October 2026, four days after the announcement:
- The Gemini API model list did not include Argon.
- The Gemini API price list did not mention Argon or Gemini 4. We searched the page for both names.
- The newest entry in the API update notes was from 22 September 2026. It was about models that turn text into speech.
That matches what Google says: paying developers cannot use Argon yet. It also means the $2 and $10 prices are only an announcement. They are not yet a price list you can be charged from.
Update, 17 October: that has changed. Google released Argon in the Gemini API on 13 October, its update notes say. When we checked on 17 October, it was on the model list and the price list. See the update below.
What does this mean if you want to use it?
If you are not in the Fairwind Program, you cannot use Argon today. And there is no date for when you can.
- For developers: you can already use Gemini 3.8 Flash. Google calls it its "most intelligent Flash model". It costs $0.75 in and $3.75 out per million tokens until 31 December 2026. If you plan a project around Argon, remember the later price. $4 and $20 is among the highest prices around.
- For Gemini app users: Google names Google AI Ultra subscribers among the first group. It does not say if people on cheaper plans or outside the US will follow, or when.
- For security teams: the Fairwind Program is the way in today. Google's Fairwind page explains the programme.
We will update this article when Argon shows up in Google's documents or price list for developers.
Update, 17 October: now in the paid Gemini API, and our own test
On 13 October 2026, Google made Argon "generally available" in the Gemini API, its update notes say. So it is no longer only for the cyber defenders in the Fairwind Program. The model list calls it "New Stable". It is not on the API's free tier: the price list says Argon "is not available on the Free tier". Google also opened it in the Gemini app for people with a Google AI Pro or Ultra plan.
The price is the starting price from the announcement, now with an end date:
- Until 30 November 2026: $2 per million tokens in, $10 out.
- From 1 December 2026: $4 in, $20 out.
We tested it. We gave Argon our three secret coding jobs: fix a money reader, split a restaurant bill from a description, and read a CSV file, a spreadsheet saved as text. Each job three times, through OpenRouter, a service that gives access to many models. Argon passed all 186 of our checks.
| Model | Checks passed | Average cost per job | Middle time per job |
|---|---|---|---|
| Gemini 4 Argon | 186/186 | 1.4 cents | 24 s |
| Gemini 3.8 Flash | 186/186 | 8.5 cents | 147 s |
| Claude Sonnet 5.5 | 186/186 | 2.6 cents | 17 s |
| Claude Opus 5.5 | 186/186 | 5.9 cents | 30 s |
| GPT-6.1 Sol | 186/186 | 1.3 cents | 22 s |
| Grok 4.7 | 182/186 | 9.5 cents | 153 s |
For the time, we sorted the nine runs and took the middle one. The surprise: Argon was cheaper per job than Gemini 3.8 Flash, even though Flash costs less per token. Counted the same way (the middle of nine), Flash used about 27,000 output tokens and Argon about 1,200. Google's own price doubles on 1 December. If our OpenRouter price doubles too, the same runs would cost about 2.7 cents per job.
Claude Opus 5.5 and Grok 4.7 were added to the table later on 17 October, from our earlier runs of the same jobs. Opus also passed every check, at about 5.9 cents per job. Grok 4.7 missed 4 checks: in two of its three tries at the money job, it read "12" and "€1.000" wrong. It was also the slowest and most expensive here.
These are small, clear jobs. They do not show how Argon does on long projects.
Update, 28 October 2026: in Google's own API, Argon is only on the paid tier. Since 26 October, new paid projects keep a copy of most requests to the Generate Content API, the usual way to send text to Gemini, and of the answers. That is what Google's changelog says. Older projects keep their setting. The copies show up as logs, a list of your past requests and answers, in Google AI Studio. Google says it does not use them to improve its products, unless you choose to share them. By default they are deleted after 55 days, Google's logs page says. You can switch them off in AI Studio's settings, or for one request.
Questions people ask
Has Gemini 4 been released?
Yes. Since 13 October 2026 Gemini 4 Argon is generally available in the paid Gemini API, and in the Gemini app for Google AI Pro and Ultra subscribers. Google announced it on 30 September 2026, but at first only Google's own teams and trusted cyber defenders in its Fairwind Program could use it. It is not on the Gemini API's free tier.
How much is 1 million tokens in Gemini?
It depends on the model. Gemini 4 Argon costs $2 per million tokens in and $10 out until 30 November 2026, and $4 and $20 from 1 December. Gemini 3.8 Flash costs $0.75 in and $3.75 out until 31 December 2026. Tokens are the small pieces of text an AI reads and writes.
How we checked
On 4 October 2026 we read Google's announcement of 30 September 2026. That included the small note with the price after the starting period. The same day we read the Gemini API model list, price list, update notes and the Gemini 3.8 Flash model page. We also read the normal price pages for developers of Anthropic, OpenAI and xAI. The prices in this article are normal list prices. They can change at any time.
We could not use Argon ourselves, because we did not have access. All claims about how well it works are Google's. That includes the test scores, the writing limit and the examples from inside Google. The screenshots are our own captures of Google's public pages. We only cropped and resized them. We made the charts and diagrams ourselves from the pages listed above. None of the images is AI-generated.
This is a summary of public sources, not a test of the model. We will add a dated note when more people can use Argon.
Update, 17 October 2026: we read Google's update notes, model list and price list again. We then ran Argon on our private secret test set through OpenRouter: three jobs, three times each, 186 checks, with the default settings. The cost is what OpenRouter charged, about 12 cents in total, paid by the owner of this site. The other models in the table ran on the same jobs between 6 and 10 October (GPT-6.1 Sol through OpenAI itself).
Correction and update, 17 October 2026: the cost of GPT-6.1 Sol was not billed by OpenRouter. It ran through OpenAI directly, so we worked out its cost from OpenAI's token counts and list prices. The other costs are what OpenRouter billed. We fixed the note under the chart and added a short animation of the same numbers, made by us from the same data.
Update, 28 October 2026: we read Google's changelog, its page on logs and datasets and its logs policy for the change in how requests are stored.
- Google: Gemini 4 Argon, our next era of frontier intelligence (30 September 2026)
- Google DeepMind: Fairwind Program
- Google: Fairwind Program announcement
- Gemini API: models
- Gemini API: pricing
- Gemini API: release notes
- Gemini API: Gemini 3.8 Flash
- Anthropic: Claude API pricing
- OpenAI: API pricing
- xAI: models and pricing