Gemini 4 Argon: what Google announced, who can use it, and what it will cost

On 30 September 2026 Google announced Gemini 4 Argon, its new frontier model. The headline numbers are big: an output limit of one million tokens, top scores on several of Google's chosen benchmarks, and an introductory price of $2 per million input tokens and $10 per million output tokens. The catch is access. For now, Argon is only rolling out to a group of trusted cyber defenders. We read the announcement, the developer documentation and the price pages of four AI companies to answer three questions: what Google announced, who can use it, and what it will cost.

What Google announced

The announcement is written by Koray Kavukcuoglu, senior vice president at Google DeepMind and Google's chief AI architect. Google calls Argon "our next era of frontier intelligence" and pitches it at three kinds of work: software engineering, knowledge work such as legal and finance, and cybersecurity defence.

Most of the post is about how Google already uses the model. Google says thousands of its employees work with it, and gives examples: memory optimisations across its data centres that freed "over 300 TiB of memory", and agents that reworked the Rust version of an open-source video decoder, which Google says now runs 2.7 times faster. These are internal results that outsiders cannot check, so we report them as Google's claims.

Before the announcement, reports about a model with the codename Argon circulated online. They disagree on when it first appeared, and we could not confirm either date, so we leave that story aside here.

Who can use it, and when

Not most people, yet. Google says Argon "is rolling out to a set of trusted cyber defenders through our Fairwind Program". Fairwind is the security programme Google launched in September 2026 for governments and companies that defend networks, with an earlier cyber model, Gemini 3.8 Flash Cyber. Google's own teams also use Argon.

Google calls this "a phased approach". It says it is taking part in "the U.S. government's voluntary process for pre-release model access" and will keep improving its safeguards with early testers first. The wider release will then be for "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. The announcement does not say when the next step happens, and it does not mention Europe or any other region. If you live in the Netherlands, as the author of this site does, there is no way to know yet whether Argon will arrive at the same time as in the US.

What it will cost

Google gives two prices, in US dollars per million tokens:

The second line comes from a footnote under the announcement, shown in the screenshot in this section. Google does not say how long the introductory period lasts.

To put that in context, we compared the standard list prices on the price pages of Anthropic, OpenAI and xAI on the same day, as in the chart at the top of this article:

At the introductory price Argon costs the same per token as Claude Sonnet 5.5 and GPT-6.1 Sol. After the introduction it matches Claude Opus 5.5 exactly. Per-token prices are only half the story: a model that thinks longer uses more output tokens for the same task, and Google is encouraging exactly that with a larger output limit. We saw that effect in our own tests: in Codex, GPT-6 Astra used about six times the credits of GPT-6.1 Sol for the same result, and Grok 4.7 cost twice as much as Grok 4.6 at the same token price.

A one-million-token output limit, and Google's benchmarks

The most concrete technical change is the output limit. Google says it is expanding the model's output limit "to an industry-leading 1M tokens, up from the previous 64K tokens". For comparison, Google's own model page for Gemini 3.8 Flash lists an output limit of 65,536 tokens, which is the 64K Google refers to. A larger limit lets a model write or reason much longer in one go, for example during a large code migration.

Google also publishes benchmark scores. Argon scores 77.9% on DeepSWE v1.1, which measures long software engineering tasks, 51.3% on AutomationBench, Zapier's test of business tasks, and 91.7% on LVBench, a test of understanding long videos. These are Google's figures from Google's chosen tests, and we could not reproduce them.

Google's own table is worth reading to the end. On FrontierSWE v2, another coding benchmark in the same table, Argon scores 55.0% and is behind GPT-6 Astra (65.5%), Claude Opus 5.5 (62.3%) and Claude Fable 5.1 (56.3%). A model can lead on one coding test and trail on another.

Cybersecurity, with fewer guardrails for some

Cybersecurity is the reason for the unusual rollout. Google says Argon can "autonomously find, validate, and patch critical software vulnerabilities". For trusted defenders and its own teams, Google writes that it will release Argon "without cyber guardrails", so that they can use its full abilities.

That explains the order. A model that is very good at finding security holes is useful to defenders and to attackers alike. Google says it is strengthening its safeguards against misuse before the broader release, including monitoring the model's reasoning for behaviour that goes beyond what the user asked.

Google also names an outside user: the security company Wiz, which Google says uses Argon in a free programme to find exposed systems in public infrastructure. On CWE-bench v1, a test of fixing security vulnerabilities, Google says Argon "ties for first place" with 68%.

Not in Google's developer documentation yet

We checked whether developers can find Argon where they would normally look. 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 text for both names.
• The newest entry in the API release notes was dated 22 September 2026, about text-to-speech models.

That fits what Google says: Argon is not yet available to paying API customers. It also means that the $2 and $10 prices are an announcement, not yet a price list you can be billed against.

What this means if you want to use it

If you are not part of the Fairwind Program, you cannot use Argon today, and there is no date for when you can.

• For developers: what you can call in the API today includes Gemini 3.8 Flash, which Google calls its "most intelligent Flash model", at $0.75 input and $3.75 output per million tokens until 31 December 2026. If you plan a project around Argon, keep the price after the introduction in mind: $4 and $20 is top-tier pricing.
• For Gemini app users: Google names Google AI Ultra subscribers among the first wave. It does not say whether 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 describes the programme.

We will update this article when Argon appears in Google's developer documentation or price list.

How we checked

We read Google's announcement of 30 September 2026, including the footnote with the price after the introductory period, on 4 October 2026. That same day we read the Gemini API model list, price list, release notes and the Gemini 3.8 Flash model page, and the standard API price pages of Anthropic, OpenAI and xAI. The prices in this article are list prices for the standard tier and can change at any time.

We could not use Argon ourselves, because it was not available to us. All performance figures, including the benchmarks, the output limit and the internal examples, are Google's claims. The screenshots in this article are our own captures of Google's public pages, cropped and resized only. The charts and diagrams were made by us from the pages listed above. None of the images is AI-generated.

This is an editorial summary of public sources, not a test of the model. We will add a dated note when Argon becomes available more widely.