Google launches Gemini 4 Argon with 1 million token limit and a focus on cybersecurity

On Wednesday, October 30, Google unveiled its new Gemini 4 Argon artificial intelligence model, designed for programming, complex professional tasks, and cybersecurity. Initial access will be limited to security partners participating in the Fairwind program, as the company finalizes protective measures before expanding its distribution.

The announcement, signed by Koray Kavukcuoglu, Senior Vice President of Google DeepMind and Chief AI Architect at Google, highlights that Argon was developed to perform prolonged, multi-phase operations. Use cases include financial research, drafting legal documents, and modifying large software projects.

The gradual rollout to developers, businesses, and consumers will begin with paying API customers and Google AI Ultra subscribers. A general release date has not yet been set.

Responses of up to 1 million tokens

One key update is the increased output limit, which has risen from 64,000 to 1 million tokens—the units of text processed by the model. According to Google, this expansion enables more extensive reasoning and the generation of broader results in a single execution.

The limit applies only to the model’s output, not to the amount of data it can receive. The company attributes this increase to solving problems that require multiple stages of work.

The introductory API pricing is set at $2 per million input tokens and $10 per million output tokens. Cached input tokens will receive a 95% discount on the standard price for that category.

Google already uses the model internally

Thousands of Google employees already use Argon for programming, research, and writing. Examples include agents who analyzed memory usage in data centers and implemented optimizations that freed up over 300 tebibytes of storage.

The model is also being used to convert C and C++ code to Rust, a language that offers advanced memory security mechanisms. The projects include libraries and over 800,000 lines of the Zircon kernel, a core component of the Fuchsia operating system.

Before reaching production systems, all changes undergo automated testing, audits, and human review.

In public tests, Argon scored 77.9% on DeepSWE v1.1, which focuses on software engineering tasks, and 51.3% on Zapier’s AutomationBench, which evaluates business process execution. Google claims the model leads these evaluations.

Cybersecurity comes first

The initial distribution prioritizes organizations working in cybersecurity. Google states that Argon can autonomously identify, validate, and fix software vulnerabilities.

A version without specific cybersecurity restrictions will be made available to select partners and internal teams, allowing full use of these capabilities in defensive activities.

Wiz already uses the model in its Scan for Good initiative, which provides free protection for critical public infrastructure. Argon detected a vulnerability that exposed personal information in healthcare software used by hospitals, though the supplier was not identified.

The launch comes amid ongoing debates about controlling these systems. During tests in May, versions of Gemini accessed systems at three real companies after an accidental internet connection. According to Google, no damage occurred.

For Argon, the company is enhancing protection against malicious instructions, monitoring actions that exceed user intent, and isolating testing environments. Google also participates in the U.S. government’s voluntary program for accessing models before launch. Feedback from early users will be used to refine these protections before wider distribution.

Frequently Asked Questions

What is Gemini 4 Argon? It is an artificial intelligence model announced by Google on September 30, 2026, focused on programming, multi-phase professional tasks, and cybersecurity.

Who can use Gemini 4 Argon? Initially, select cybersecurity partners from the Fairwind program. Google plans to expand access, starting with paying API customers and Google AI Ultra subscribers, but has not yet set a date.

How much will Gemini 4 Argon cost? The introductory API pricing is $2 per million input tokens and $10 per million output tokens. Cached input tokens will receive a 95% discount.

What does the 1 million token limit mean? This is the model’s output limit, designed to enable more extensive executions and results. It does not refer to the amount of information the system can receive.