Gemini 4 Argon Goes Global: 1M Output Tokens and a Cyber Focus

Google took Gemini 4 Argon global on October 13, 2026. Before that, the model was shared on a limited basis through the Fairwind Program, an early-access track for selected cyber defense teams. Argon is now available through a paid API and the Google AI Ultra subscription.

Argon was first announced on September 30, 2026. What is new on October 13 is the reach: no longer a small circle of testers, but paying customers in many countries. It marks Google's effort to reclaim ground that OpenAI and Anthropic had taken.

What stands out most about Argon is not its scores, but one number: a 1 million token output limit. Earlier Gemini models were capped at 64,000 output tokens. That means Argon can write a reply roughly fifteen times longer, enough to produce a large document in a single pass.

Why a Long Output Matters

An output limit is not just about writing long text. It decides which kinds of work are possible.

A model with a short output must break a large task into many chunks. Each chunk risks losing context, and the results must be stitched back together. A model with a 1 million token output can handle a large task in one breath, from reading code to writing its migration.

Google gives a real example from inside the company. Its quantum computing research team used Argon to optimize the spacetime resources of subroutines that bottleneck important applications. In one case, Argon beat a published baseline by 40 percent, achieved in minutes. That claim comes from Google itself, so read it as a claim by the model's maker.

The Numbers Google Published

Google published several benchmark scores, and these are what make Argon comparable.

Google claims Argon leads the Vals Index, a benchmark that measures economic impact in finance, coding, legal, and tax work. The "leads" claim needs independent verification, since rivals also use the Vals Index. But the direction is clear: Google positions Argon not as a chatbot, but as a professional work tool.

Cybersecurity: The Unusual Differentiator

The most unusual part of Argon is its focus on cyber defense. Google trained Argon to find, validate, and patch software vulnerabilities on its own.

For trusted testers and Google's internal teams, Argon is released without cyber guardrails, so its full capability can be used. Wiz, a security company, uses Argon through its Scan for Good program. In an early demonstration, the model found a critical vulnerability exposing personal information across hospital software worldwide, a risk that previous frontier models had missed.

Google also reports that Argon agents are working on migrating C/C++ codebases to Rust inside Google, from tens of thousands of lines in core libraries up to more than 800,000 lines for the Fuchsia Zircon kernel. For libgav1, Argon agents replaced 32,000 lines of SIMD code, and the result is a video decoder 2.7 times faster than the earlier Rust port. These figures come from Google, and such large migrations still go through strict auditing before reaching production.

Fairwind: Why the Rollout Was Gradual

The gradual launch of Argon is not accidental. Google says Argon first went to "trusted cyber defenders" through the Fairwind Program before reaching the public. A model that can find and patch software vulnerabilities is a powerful tool, and the same tool can be used for the opposite.

Google is also taking part in the U.S. government's voluntary process for pre-release model access. That means before Argon was broadly available, the government had access to test it. The step shows that big labs now accept that a model with offensive capability cannot be shipped like an ordinary product.

What is striking is that part of that capability is sold as a feature, not hidden. For trusted testers, Argon is released with no cyber guardrails at all, so its full power can be used to defend. Google places great trust in one assumption: that the right testers will use this power to protect, not to attack.

Pricing and How to Get It

Argon is priced at an introductory $2 per million input tokens and $10 per million output tokens. After the introductory period, the price rises to $4 per million input and $20 per million output. Cached input tokens get a 95 percent discount.

For comparison, Anthropic's Claude Opus 5.5 is priced at $4 input and $20 output per million tokens. OpenAI's GPT-6 Astra is $10 input and $50 output. That means at the introductory price, Argon is cheaper than both; at the regular price, it matches Opus on the input side.

Safeguards Cited Before Broad Release

Google names four areas of protection before Argon was made broadly available. First, defense against misuse for cyberattacks and weapons of mass destruction, including monitoring the model's internal activations. Second, resilience against prompt injection, an attack that slips malicious instructions into the context. Google claims Argon leads on the Gray Swan benchmark.

Third, misalignment monitoring, meaning attempts by the model to deviate from the user's intent. Google monitors Argon's chain of thought and actions, then stops execution when necessary. Google also warns that findings from monitoring should not be fed back into training, so the model does not learn to evade monitoring. Fourth, hardening the sandbox environments where the model is tested.

Analysis: Why Google Put Its Chip on Cyber Defense

Google's strategy with Argon is interesting because it is not competing on the same ground as its rivals. OpenAI and Anthropic dominate the narrative of do-everything assistants and general agents. Google picked a specific ground: professional work and cyber defense.

There is a practical reason. A model that can find software vulnerabilities has clear selling value to large companies, and that value is not easy to compare through chatbot benchmarks. Three things follow from that choice.

• A clearer buyer. Security teams and large enterprises have a budget and a specific need.
• A harder-to-copy claim. Vulnerability discovery is measured by results on real systems, not by a chat demo.
• A narrower market. The same focus makes Argon less of a general consumer product.

This shifts the question from "which model is smartest" to "which model can be trusted with high-stakes work".

This focus also explains why Google released Argon gradually. A model that can attack and defend networks must be handled more carefully than a model that only writes poetry. Sending it to cyber defense teams first, then to the public, is a way to test capability while controlling risk.

But there is an unresolved tension. Google releases Argon without cyber guardrails to trusted testers. The same model that can patch a vulnerability can also be used to find one to attack. Google places great trust in the line separating the two, and that line is not always clear.

What This Means for Users in Indonesia

For users in Indonesia, three practical points follow from this release.

• The 1 million token output limit opens new work. Summarizing a legal document hundreds of pages long, migrating a large codebase, or producing a long report can now be done in one step, instead of being split into dozens of chunks stitched by hand.
• The price per token decides who can use it. At the regular $4 and $20, Argon matches Opus. For developers on a tight budget, the difference in the introductory $2 and $10 may matter at scale.
• The cyber focus opens a local opportunity. Security teams in Indonesia could use the model for vulnerability audits, but the results still need human verification.

One caveat: most performance claims come from Google itself, and benchmarks do not always reflect real tasks. Before moving important work to Argon, testing on your own data is more useful than a benchmark score.

An unanswered question: if the strongest models are now marketed for high-stakes professional work, who verifies the answer is correct when the model is wrong?