Google Hands Gemini 4 Argon to Cyber Defenders First, With a 1M-Token Output Limit

Gemini 4 Argon: Google's Strongest Model Reaches Defenders Before Everyone Else

On September 30, 2026, Google DeepMind announced Gemini 4 Argon, its new frontier model for software engineering, knowledge work and cyber defense. According to Google, the model can autonomously find, validate and patch critical software vulnerabilities, and it raises the limit for a single response to 1 million tokens, up from 64,000. It is not generally available yet: Google is rolling it out first to trusted cyber defenders through its Fairwind program, then to paid API customers and Google AI Ultra subscribers. The introductory price is $2 per million input tokens and $10 per million output tokens.

The news here is not another new model. It is that Google chose to hand its strongest model to the people who protect systems before it reaches the people who might attack them. That decision alone tells you where security risk is heading over the next year.

What Google Announced, in Numbers

• Output limit: 1 million tokens in a single response, up from the previous limit of 64,000 tokens.
• Introductory price: $2 per million input tokens and $10 per million output tokens, with cached input 95% off.
• Price after the introductory period: $4 for input and $20 for output per million tokens. Build your budget on this number, not the launch price.
• Published benchmark results: 77.9% on DeepSWE v1.1 for software engineering, 51.3% on AutomationBench where it ranks first, 68% on CWE-bench v1 for vulnerability work where it is tied for first, and 91.7% on LVBench for long-video understanding.

These are Google's own figures. Independent testing takes weeks, and early third-party results are mixed, which is a reminder to test any model on your own tasks before moving real work onto it.

Why Defenders First? The Fairwind Program

Fairwind is Google's way of giving high-priority defenders, such as government bodies, healthcare providers and telecommunications companies, early access to advanced models, so they get a head start on attackers before a model becomes available to everyone. Google also says it is taking part in the U.S. government's voluntary process for pre-release model access, and that Argon went through testing by internal and external red teams, with safeguards that monitor the model's chain of thought and its actions.

The logic is simple: a model that can find a vulnerability and patch it can, in the wrong hands, find it and leave it open. Giving defenders an early start shrinks the window in which the attacker is ahead.

What Google Says Argon Achieved Internally

Google reports that Argon helped free more than 300 TiB of memory across its systems, and helped migrate more than 800,000 lines of the Fuchsia Zircon kernel to Rust. These are the results of a giant company on its own infrastructure, but they show the kind of work the model was built for: long, multi-step engineering tasks, not short chat answers.

What a 1-Million-Token Limit Changes in Practice

Most models stop writing before they finish a large task. A 1-million-token limit means the model can, in one pass, produce a complete migration of an old module in your system, a full technical report, or a thorough security review of a codebase. That is useful, and also expensive if you are not careful: one response that uses the full million tokens costs $10 at the introductory price and $20 at the standard price. Multiply that by the number of attempts and the number of employees, and the bill grows quickly. The real metric is cost per completed task, as we explained in our article on AI cost per task.

The Security Lesson Does Not Wait for Access

You cannot use Argon today, and that changes little. The direction is clear: AI that finds vulnerabilities on its own is coming, and on both sides. Old, unpatched systems will be discovered faster than before. We saw this pattern in the WordPress attacks that began within hours of a fix being published and in network equipment whose vulnerabilities are being actively exploited. Steps worth starting now:

• Know what you run. Keep an up-to-date list of every system, plugin and library your business relies on, and who is responsible for updating each one.
• Shorten your patch time. If applying a critical update takes you weeks, the gap is now longer than the time an attacker needs.
• Ask your software provider directly. Who monitors vulnerabilities in your system, and how long does it take to ship a fix? Then write the answer into the contract.
• Check yourself against national controls. The Essential Cybersecurity Controls issued by the National Cybersecurity Authority are a practical baseline for organizations in Saudi Arabia.

Before You Adopt Any New Model

When Argon reaches the paid API, the question will come up: should we switch? Our advice is the same with every launch: do not tie your systems to a single provider. Build an intermediate layer that lets you compare Gemini, Claude and GPT on the same task and move between them as price or quality changes, as we discussed when comparing GPT-6 Astra and Claude Fable 5.1. And if personal data about your customers or employees will pass through the model, check where it is processed and whether its transfer complies with the Personal Data Protection Law.

At Origami we build systems that connect AI to your business data through a layer you control: you choose the model, define what it can see, and measure the cost of every task. When a stronger model arrives, switching to it becomes a configuration change, not a rebuild.

Sources

• Google's official announcement: Gemini 4 Argon (September 30, 2026)
• National Cybersecurity Authority: Essential Cybersecurity Controls
• SDAIA: Personal Data Protection Law