Gemini 4 Argon Explained: Google’s New AI Model, Price, and When You Can Try It

On September 30, 2026, Google announced Gemini 4 Argon, the first model in its new Gemini 4 generation and, in the company’s words, its most powerful AI system to date. It is built for long, complex work: software engineering, enterprise knowledge tasks like legal and finance, and defensive cybersecurity. But you cannot use it yet. Google is releasing it in phases, starting with a small group of vetted cyber defenders, with no public release date announced.

This guide breaks down what Argon actually is, what it costs, how it stacks up against OpenAI’s GPT-6 and Anthropic’s Claude models, and when ordinary users might get access.

What is Gemini 4 Argon?

Gemini 4 Argon is a frontier AI model from Google DeepMind, announced in an official Google blog post on September 30, 2026. According to Google, it is designed to “sustain deep reasoning across complex, long-horizon workflows,” which is the company’s way of saying it can work through multi-step tasks that take a long time, like migrating a large codebase, analyzing lengthy legal documents, or hunting down software vulnerabilities.

The name follows Google’s pattern of giving its top-tier models codenames. Argon sits above the lighter-weight Gemini Flash models and marks a shift back toward large frontier models, after the company had spent the previous generation focusing on smaller, faster, high-efficiency systems.

What it is built for

Google is positioning Argon around three main workloads:

• Cybersecurity defense. This is the headline use case. Google says the model was trained specifically for defensive security work and can autonomously find, validate, and patch critical software vulnerabilities. Cybersecurity is also the reason for the gated rollout: capabilities that can analyze vulnerabilities need careful handling before being handed to everyone.
• Real-world software engineering. Debugging, codebase migrations, and long-running development tasks. Google says thousands of its own employees are already using Argon internally for coding, research, and writing.
• Enterprise knowledge work. Legal and finance workflows in particular. On one legal benchmark, Google claims Argon scored far ahead of both OpenAI’s and Anthropic’s flagship models, though these are company-reported numbers and should be read with that in mind.

The standout technical change is the output limit. Argon can generate up to 1 million tokens in a single task, up from 64,000 tokens in Google’s previous advanced models. That is roughly the difference between writing a short report and writing an entire book in one sitting, and it matters for tasks where the model needs to hold an enormous amount of context without losing the thread.

How it compares to GPT-6 and Claude

Benchmark claims from AI companies always deserve a grain of salt, since each lab highlights the tests where it wins. Here is what Google and independent coverage report, with that caveat:

• Google says Argon beats OpenAI’s GPT-6 Astra, Claude 5.1, and Claude Opus 5.5 on knowledge work, science and math, and multimodal understanding, and matches on cybersecurity. It trails in coding benchmarks such as FrontierSW, TerminalBench, and OSWorld.
• On the AA-Omniscience benchmark, which measures how often a model guesses wrong when it does not know an answer, Argon posted the lowest error rate among top models at 15 percent, meaning it is unusually good at saying “I do not know” instead of hallucinating. But raw accuracy on the same test was 50 percent, behind GPT-6 Astra’s 63 percent. Honesty, in other words, not omniscience.
• On Harvey’s Legal Agent Benchmark, Argon scored 19.6 percent against GPT-6 Astra’s 5.4 percent and Claude Opus 5.5’s 3.8 percent, suggesting real strength in legal workflows.

Context matters: the launch week was crowded. OpenAI’s DevDay announcements, including its new Dot personal agents, landed the same day, and both labs have been trading the “best model” crown for months. If you are deciding between assistants for everyday work rather than API models, our Claude vs ChatGPT comparison for small business covers the two most popular consumer options right now.

Gemini 4 Argon pricing

Google published introductory API pricing with the announcement:

• Input: $2 per million tokens
• Output: $10 per million tokens
• Cached input: 95 percent off the input token price

That puts it in the same price band as OpenAI’s GPT-6 models and roughly half the cost of Anthropic’s top-tier Claude Opus pricing, according to coverage of the launch. These are introductory rates for API developers; there is no consumer subscription price yet because there is no consumer product yet.

When can you actually use it?

Not yet, and Google is not saying exactly when. The rollout is phased:

1. Now: trusted cybersecurity defenders through Google’s Fairwind Program.
2. Next: broader access for developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers.
3. Eventually: general availability, with no date given.

Google is also participating in the US government’s voluntary pre-release evaluation process, with cybersecurity experts vetting the model before wider release. The company says it is deploying “misalignment mitigations” that monitor the model’s reasoning and can shut down actions when needed, along with extra precautions around its training sandboxes to prevent agent escape scenarios.

If you are a developer, there is nothing to sign up for publicly right now. The practical move is to watch the official Gemini API documentation and Google’s AI Ultra plan for the first sign of availability. For consumer context on where Google’s AI currently shows up in products, see our piece on the AI to superintelligence renaming and what it signals about where the industry is heading.

Why the cautious rollout matters

A year ago, a frontier model launch meant a public demo and an API key for everyone. Argon’s phased release reflects a real shift: AI leaders, including OpenAI and Google, have started talking openly about slowing the frontier to let safety measures catch up. OpenAI shelved a model the same week over safety concerns, and Google chose to gate Argon behind vetted defenders first.

Whether you read that as responsible caution or marketing, the pattern is worth watching. The most capable new models are increasingly launching to enterprises and specialists first, with consumer access trailing months behind. Budget for that lag if you are planning to build on new models.

Frequently asked questions

Is Gemini 4 Argon available to the public?

No. As of October 1, 2026, it is only available to a select group of cybersecurity organizations in Google’s Fairwind Program. Google has not announced a general release date.

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 at 95 percent off. There is no consumer pricing yet.

Is Gemini 4 Argon better than GPT-6?

It depends on the task. Google’s benchmarks show Argon leading in knowledge work, science, math, and legal workflows, while trailing GPT-6 Astra in coding benchmarks and raw accuracy on some tests. All of these numbers come from the labs themselves, so treat them as marketing-adjacent until independent evaluations land.

What is the Fairwind Program?

Google’s program for rolling out advanced AI capabilities to trusted cybersecurity partners first, so defensive uses get battle-tested before wider release.

Will Gemini 4 Argon come to Google AI Ultra?

Google says paid API customers and Google AI Ultra subscribers will be among the first to get broader access, but it has given no timeline.

What does “1 million token output” actually mean?

It is the amount of text the model can generate in a single response session. One million tokens is roughly 750,000 words, enough to sustain very long, multi-step tasks without the model forgetting earlier context. In practice, most users will use a fraction of that.

Bottom line: Gemini 4 Argon looks like a genuinely significant model, especially for cybersecurity and legal knowledge work, but it remains an announcement rather than a product. The pricing is competitive, the benchmarks are promising but self-reported, and the release date is unknown. When Google opens the gates, we will update this guide.