Google Unveils Gemini 4 Argon With 1 Million-Token Output — But You Can’t Use It Yet

Google has unveiled Gemini 4 Argon, a new frontier artificial intelligence model designed for demanding tasks across software engineering, cybersecurity, finance, legal work and other complex professional workflows.

The announcement marks the beginning of Google’s Gemini 4 generation, but there is an important catch: Gemini 4 Argon is not yet broadly available to the public.

Google is initially providing the model to a limited group of trusted cybersecurity defenders through its Fairwind Program while it continues safety testing and gathers feedback before expanding access.

The company says paid API customers and Google AI Ultra subscribers are expected to be among the first broader groups to receive access, although Google has not announced a specific public release date.

What Is Gemini 4 Argon?

Gemini 4 Argon is Google’s new high-end AI model built specifically for long, complicated workflows rather than only short question-and-answer interactions.

Google says the model can maintain reasoning across extended tasks involving multiple steps, documents, tools and decisions.

Its target workloads include software development, financial research, legal analysis, cybersecurity defense, multimodal analysis and creative writing.

That positioning suggests Google increasingly sees advanced AI models not simply as chatbots, but as systems capable of performing larger pieces of professional work.

The Biggest Number: 1 Million Output Tokens

One of Gemini 4 Argon’s most unusual specifications is its dramatically expanded output limit.

Google says Argon can generate up to 1 million output tokens, compared with a previous limit of 64,000 tokens referenced in the company’s announcement.

This is important because an output limit determines how much information a model can generate during a single trajectory.

Rather than stopping after a relatively short response, Argon can potentially continue working through extremely long tasks involving extensive reasoning, code generation, analysis or document creation.

Google says the additional headroom is intended to allow the model to reason more deeply on difficult problems.

However, a larger output limit does not automatically mean every response will be more accurate. Performance still depends on the task, prompts, tools, data and model reliability.

Gemini 4 Argon Is Already Working Inside Google

Although most users cannot access Argon yet, Google says the model is already being used internally.

One particularly interesting example involves Google’s data centres.

According to the company, a group of Argon agents analysed fleet-wide profiling information and autonomously identified memory optimisations. Google says changes already rolled out have freed more than 300 TiB of memory, with estimated potential savings of between 500 TiB and 1 PiB.

That example illustrates where the AI industry may be heading.

Instead of asking an AI assistant a question and receiving text back, companies are developing AI agents that can analyse complex environments, propose actions and complete portions of larger workflows.

Argon Is Helping Rewrite Google Code

Google also says Argon agents are being used for large-scale code migrations.

One project involves converting C and C++ code to the memory-safe Rust programming language.

According to Google’s announcement, the work ranges from libraries containing tens of thousands of lines to more than 800,000 lines associated with the Fuchsia Zircon kernel.

The company stressed that critical code migrations still go through automated and manual auditing, testing and review before being deployed.

In another example, Google says Argon agents worked on its open-source libgav1 video decoder and replaced around 32,000 lines of SIMD code in an existing Rust port.

Google reports that the resulting decoder ran 2.7 times faster than that Rust port while maintaining identical video output.

These are company-reported results, so independent real-world testing will become important once developers receive broader access.

Coding Is One of Gemini 4 Argon’s Main Targets

Google is heavily positioning Argon toward professional software engineering rather than basic code completion.

The company reports a score of 77.9% on DeepSWE v1.1, a benchmark designed around long-horizon real-world software engineering tasks.

But benchmark comparisons need context.

Reuters noted that although Google’s published results showed Argon ahead of competing frontier models on several evaluations, it remained behind rivals on some other metrics, including two of the four coding-related benchmarks included by Google.

That means it would be premature to describe Argon as universally superior based only on launch-day benchmark numbers.

Independent testing will provide a clearer picture once broader access begins.

Google Is Pushing Argon Into Finance and Legal Work

Coding is only one part of Google’s strategy.

Gemini 4 Argon is also designed for complex enterprise knowledge work.

Google says the model performed strongly on evaluations involving financial research, legal research and drafting, tax work and other professional activities.

On AutomationBench, a benchmark from Zapier that evaluates end-to-end execution across business functions, Google reports that Argon achieved a score of 51.3%.

The significance here is not simply whether AI can answer financial or legal questions.

The bigger shift is toward enterprise AI agents that can work through multi-stage business processes involving different documents, software systems and actions.

That could eventually make models such as Argon relevant to companies building automated workflows across finance, operations, software development and other departments.

Cybersecurity Is Why Google Is Being Careful

Cybersecurity appears to be one of Gemini 4 Argon’s most powerful—and potentially sensitive—capabilities.

Google says Argon can autonomously find, validate and patch critical software vulnerabilities.

That creates an unusual challenge.

The same reasoning abilities that can help security teams discover weaknesses could potentially be misused to search for exploitable vulnerabilities.

For this reason, Google is not immediately providing unrestricted Argon access to everyone.

Trusted cyber defenders and Google’s internal teams are receiving specialised access while broader safeguards continue to be tested.

Argon Found a Vulnerability Previous Models Missed

Google says cloud-security company Wiz has been testing Argon through its Scan for Good initiative.

During an early deployment, Google reports that Argon discovered a critical vulnerability affecting healthcare software used by hospitals around the world.

According to Google, the vulnerability could expose sensitive personal information and had been missed by previous frontier models.

Google also reports that Argon tied for first place with a 68% score on CWE-bench v1, which measures AI systems’ ability to remediate software security vulnerabilities.

Again, these launch claims will become easier to assess as independent researchers gain access to the model.

Google Is Testing Against Prompt Injection

Another major security challenge for AI agents is prompt injection.

A prompt-injection attack can place malicious instructions inside content that an AI system reads, potentially attempting to manipulate the model into performing actions the user did not intend.

This becomes especially important as AI models gain access to email, files, browsers, enterprise applications and other tools.

Google says Gemini 4 Argon has undergone automated red teaming and adversarial training designed to improve its resistance to indirect prompt-injection attacks.

The company is also developing monitoring systems intended to detect when an AI agent attempts to move beyond the user’s intended task.

Why Can’t Everyone Use Gemini 4 Argon Yet?

The limited launch is deliberate.

Google says it is continuing safety work in four major areas before making Argon broadly available, including misuse prevention, prompt-injection protection, monitoring for unintended behaviour and stronger sandboxed environments for testing advanced AI systems.

The company is also participating in a US government voluntary process involving pre-release access to advanced models.

So users searching for a Gemini 4 Argon download, public API or immediate Gemini app access should be careful.

There is currently no general public release.

Google also has not announced a fixed date for when everyone will receive access.

Who Will Get Gemini 4 Argon First?

The rollout is expected to happen gradually.

Right now, trusted cybersecurity defenders participating in Google’s Fairwind Program are among the early users.

Google says broader availability will eventually extend to developers, enterprises and consumers, with paid API customers and Google AI Ultra subscribers expected to be among the first groups in that expansion.

That wording is important because it does not mean every AI Ultra subscriber already has Argon today.

The broader rollout is still upcoming.

Gemini 4 Argon Pricing Revealed

Google has already disclosed introductory API pricing even though widespread API access has not begun.

The introductory price is planned at:

Usage	Introductory Price
1 million input tokens	$2
1 million output tokens	$10
Cached input	95% discount on input price

After the introductory period, Google says pricing will increase to $4 per million input tokens and $20 per million output tokens.

This pricing is particularly relevant for developers and companies considering Argon for long-running AI-agent workflows, where token consumption can become substantial.

What Happened to Gemini 3.5 Pro?

Gemini 4 Argon also represents a change in Google’s previously expected model roadmap.

Reuters reports that Google no longer plans to release Gemini 3.5 Pro, which had previously been expected earlier in 2026.

Instead, Google has moved forward with Gemini 4 and Argon as its new top-tier model.

That makes Argon more than another incremental Gemini update—it effectively becomes the company’s next major frontier-model generation.

Can You Use Gemini 4 Argon in India?

Not generally yet.

The announcement is global, but Google has not opened Argon to ordinary Gemini users in India or other markets at this stage.

Indian developers therefore should not confuse the announcement with immediate availability through the Gemini API.

The first broader rollout is expected to target paid API customers and Google AI Ultra subscribers, but Google has not provided a country-by-country schedule.

Users should rely on Google’s official Gemini 4 Argon announcement for rollout information rather than unofficial websites claiming to provide early access.

Why Gemini 4 Argon Matters

The most important development may not be one benchmark score.

Argon shows how leading AI systems are moving from generating short answers toward sustained autonomous work.

A future enterprise AI system might analyse thousands of files, inspect a software repository, identify a problem, use approved tools, make changes, test the result and prepare a final report.

That is a much more complicated workload than today’s familiar chatbot interaction.

It is also why security becomes more important.

The more authority an AI agent receives, the greater the potential impact when it makes an error or follows malicious instructions.

Google’s cautious rollout of Argon reflects that tension: advanced agent capabilities could create major productivity gains, while simultaneously requiring stronger controls before they are widely deployed.

What Happens Next?

The immediate milestone to watch is broader access.

Google says Gemini 4 Argon will eventually roll out to developers, enterprises and consumers, beginning with paid API customers and Google AI Ultra subscribers, but there is no confirmed general-public release date yet.

Until independent developers can test it at scale, Google’s benchmark numbers and internal productivity examples should be treated as company-reported results rather than a final verdict on how Argon compares across every real-world workload.

For now, the confirmed picture is clear: Gemini 4 Argon has been announced, it supports up to 1 million output tokens, Google is already using it internally, cybersecurity defenders are receiving early access—and most users still have to wait.