Google's Gemini 4 Argon is here, but only cyber defenders can use it so far

On September 30, 2026, Google announced Gemini 4 Argon, which it calls "our new frontier model." Most people still can't use it. Google is releasing Argon first to "a set of trusted cyber defenders" through a Google DeepMind access scheme called the Fairwind Program. Google says those defenders get the model "without cyber guardrails." Paid API customers and Google AI Ultra subscribers are next in line, but Google gave no date. As of October 9, Argon does not appear on the Gemini API's public model list or pricing page.

That makes Argon an unusual launch. The price, the benchmark claims and the safety argument are all public, but the product is not.

What Google announced

The launch post is written by Koray Kavukcuoglu, Google DeepMind's SVP and Google's chief AI architect. It describes Argon as a model built to "sustain deep reasoning across complex, long-horizon workflows." Google points it at three kinds of work: software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense.

Confirmed details from Google's post:

• Access: First to vetted defenders through Fairwind, and to internal Google teams. Google says it is taking "a phased approach" and is taking part in the US government's voluntary pre-release model access process.
• Next in line: Google says Argon is "rolling out soon," "starting with paid API customers and Google AI Ultra subscribers." The post doesn't say when, or mention the free Gemini app.
• Price: An introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After that, $4 input and $20 output. Google doesn't say how long the introductory period lasts.
• Output length: Google says it is raising the output token limit to 1 million tokens, up from 64K.

One thing Google has not published is Argon's context window, meaning how much input the model can read at once. It isn't in the launch post, on the DeepMind model page, or in the evaluation methodology document. Some third-party sites list a figure. We haven't found one from Google, so we don't repeat any.

Last verified: October 9, 2026

Who gets in: the Fairwind Program

According to DeepMind's program page, Fairwind is for governments and national cyber authorities, critical infrastructure operators (healthcare, telecom, energy, finance), core technology platforms, and academic labs doing defensive benchmarking. The terms are strict:

• Applicants go through background checks.
• Access requires phishing-resistant multi-factor authentication.
• Use is limited to internal security, incident-response or penetration-testing teams.
• Partners may not share, redistribute or sell access.
• Malware creation is explicitly banned.

Google also says Argon "supports zero data retention" on Gemini Enterprise for these partners.

The same program already gates Gemini 3.8 Flash Cyber, a security-tuned model Google released on September 2. Google's reasoning is that the skills that let a model "autonomously find, validate, and patch critical software vulnerabilities" can be turned to attack just as easily. Giving defenders a head start is the point.

What Google claims Argon can do

These are Google's claims. Independent evaluations of Argon have not been published, and outside access is limited to Fairwind partners.

Google DeepMind's model page compares Argon with OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5 on 19 benchmarks. On Google's numbers, Argon has the top score, or ties for it, on 14:

• DeepSWE v1.1, long-horizon coding: 77.9%, versus 74.2% for Opus 5.5 and 74.1% for GPT-6 Astra.
• Harvey's Legal Agent Benchmark: 19.6%, versus 6.7% or less for the others.
• GraphWalks at 256,000 to 1 million tokens, a long-context test: 84.2%, versus 71.8% for GPT-6 Astra.
• CWE-bench v1, a cybersecurity benchmark: 68.0%, tied with GPT-6 Astra.

By the same table, Argon trails on five benchmarks. Claude Opus 5.5 leads Terminal-bench 4.0 (66.4% vs 57.4%) and PostTrainBench. GPT-6 Astra leads FrontierSWE v2 (65.5% vs 55.0%), Terminal-Bench Science and the OSWorld-2.0 computer-use subset.

Google's methodology document flags several caveats. Most competitor scores are reported by the vendors themselves or taken from public leaderboards, and settings differ from model to model. The OSWorld figure is the best of three runs rather than one attempt. Terminal-Bench Science used a verifier timeout six times the default. Video frame sampling also differed across models on LVBench. Treat the table as Google's case for Argon, not a neutral ranking. OpenAI's GPT-6 lineup is covered in our GPT-6 report, and Anthropic's Opus 5.5 and Fable 5.1 in which Claude model to use.

The post also cites internal results that haven't been independently verified. Google says Argon beat a published quantum-computing baseline by 40%, freed more than 300 TiB of memory across Google's server fleet, and is migrating C and C++ codebases of up to 800,000+ lines to Rust. It also says the security firm Wiz, through its Scan for Good initiative, used Argon to uncover "a critical vulnerability exposing sensitive personal information across healthcare software."

What developers can use today

If you build on the Gemini API, nothing changed on September 30. The current lineup on Google's models page, last updated October 6, 2026:

Last verified: October 9, 2026 (ai.google.dev models and pricing pages)

Some coverage refers to a "Gemini 3.5 Pro." Google's models page lists no such model. The only Pro-class text model in the API is Gemini 3.1 Pro, and it is still a preview. Google also lists 3.7 Flash and 3.6 Flash as stable, with 3.5 Flash marked legacy. On September 18, Google limited Gemini 2.5 access to users who had used those models recently.

Nothing newer than Argon shipped between October 1 and 8. The only Gemini API launch in that window was an image model, Nano Banana 2.1, which became generally available on October 6.

Why it matters

Argon's launch is the second time in a month that Google has kept a frontier model to vetted partners first. Gemini 3.8 Flash Cyber came out the same way on September 2. Google is treating offensive-cyber capability as the main risk in a frontier release, ahead of the usual "is it smarter" story.

For everyone else, the practical points are:

• Developers: The $4/$20 standard price is double the introductory rate. Plan budgets on the standard price, not the launch discount. Until Argon shows up on the pricing page, Gemini 3.8 Flash is the model to build on.
• Gemini app users: Ultra subscribers are promised early access, but no date has been given. Our ChatGPT vs Gemini for research comparison covers what each Google AI plan includes today.
• Long-document users: Google's long-context scores and the 1M-token output limit are the headline numbers to watch. A bigger output limit says nothing about how well a model recalls what it read; our explainer on long-context recall explains why. For what Gemini can handle today, see how Gemini handles very large PDFs.

What's confirmed and what isn't

• Confirmed: Announcement date, Fairwind-first access, the announced prices, the 1M output limit, and that Argon is absent from the public API model list and pricing page as of October 9.
• Vendor-reported: Every benchmark number and internal-use claim. Wiz's finding of a critical vulnerability in healthcare software is also Google's description.
• Unknown: The context window, the general-availability date, how long the introductory price lasts, and whether free Gemini app users will get Argon.