Google Dropped Gemini 4 Argon: Here’s What It Can Do (and Who Can Actually Use It)

Ask any AI model to rewrite a large codebase and you’ll hit the same wall: it stops halfway through a file, and you spend the next hour stitching fragments together. Output limits, not intelligence, have been the quiet bottleneck for serious work.

Google announced Gemini 4 Argon on September 30, 2026, and it goes straight at that problem. Argon can write up to 1 million tokens in a single response, up from 64,000 on Gemini 3.1 Pro Preview. Google pitches it as a frontier model for coding, enterprise knowledge work, and cyber defense.

There’s a catch. You probably can’t use it yet.

Here’s what Argon does, where it beats GPT-6 Astra and Claude, where it doesn’t, and when you might get access.

Can you use Gemini 4 Argon right now?

Not unless you work in cyber defense. As of October 14, 2026, Google has not opened Argon to the public. Access follows three stages:

• Stage 1 (now): Vetted security teams in Google’s Fairwind Program, using Argon through Gemini Enterprise.
• Stage 2: Paid Gemini API customers and Google AI Ultra subscribers.
• Stage 3: Broader availability for developers, businesses, and consumers.

Google hasn’t given a date for Stages 2 or 3. It says access will widen after more feedback and guardrail work. Argon doesn’t appear on any Gemini app plan today. If you’re weighing an upgrade, check our breakdown of Gemini AI pricing and the free vs paid plans before paying for Ultra on Argon’s promise alone.

Google is also taking part in a voluntary U.S. government program that evaluates advanced models before release. That partly explains the slow rollout.

Why Gemini 4 Argon’s 1 million token output matters

Most models cap a single answer at tens of thousands of tokens. Argon’s ceiling is 1 million. The Batch from DeepLearning.AI reports that Google gets there with a technique called Long Decode Continuation. It pauses a response and resumes it across API calls, so long jobs don’t time out.

In practice, that changes what you can hand off. Instead of asking for one function at a time, you can ask Argon to port an entire module, draft a full contract set, or produce a complete audit report in one pass.

There’s a caveat: not every endpoint may serve the full million at launch. Vals AI tested Argon with a 262,000-token output cap, so real-world limits may vary by setup.

On the input side, The Batch lists text, image, and video inputs up to 1 million tokens. Google hasn’t published that as a formal spec, so treat it as reported, not confirmed.

Where Gemini 4 Argon beats GPT-6 Astra and Claude (and where it doesn’t)

Google published a 19-benchmark comparison table. Argon leads 13, ties one, and trails five. The pattern is clear once you group them.

Google reported these scores, and LLM Stats compiled them. Independent verification is still limited.

Argon wins on knowledge work: finance, legal, business automation, and long-context reasoning. It loses on the hardest agentic coding tests. FrontierSWE and Terminal-Bench both favor OpenAI and Anthropic.

So if you write code for a living, Argon isn’t an automatic upgrade. If you analyze filings, contracts, or long video, it likely is.

Gemini 4 Argon admits when it doesn’t know

Most launch coverage skipped this stat, and it’s the one I’d watch.

Artificial Analysis runs a test called AA-Omniscience that penalizes wrong answers but not refusals. Argon posted a 15% hallucination rate on it. That’s the lowest among models scoring 45 or higher on the Intelligence Index. GPT-6 Astra at maximum reasoning scored 51% on the same test.

Why does that matter more than a benchmark win? A refusal is obvious. A confident wrong answer in a tax memo or legal brief can slip past a reviewer and cost real money. For finance and law teams, a model that says “I don’t know” is often worth more than one that scores two points higher.

The tradeoff: you’ll need a plan for the questions Argon declines.

What Google already built with Argon internally

Thousands of Google employees already use Argon. The internal results show what long output makes possible:

• Rust migration at scale. Agents translated over 800,000 lines of the Zircon kernel (the core of Google’s Fuchsia OS) plus other C and C++ code into Rust.
• A faster video decoder. Google reworked about 32,000 lines of SIMD code in the libgav1 decoder. The Rust version runs 2.7 times faster than the previous port with identical output.
• Freed data center memory. Agents analyzing data center operations found ways to free over 300 TiB of RAM. Google projects 500 TiB to 1 PiB after a wider rollout.

The Rust work is the quiet headline. Memory-safety bugs cause a large share of security flaws in C and C++ code. A model that can rewrite whole codebases in a memory-safe language lets teams remove an entire class of bugs instead of patching them one by one.

Why Google locked Argon behind a cyber-defense program

Argon is trained to find, verify, and fix software vulnerabilities. That’s useful for defenders and dangerous in the wrong hands, which is why the first users are security teams.

Wiz, the cloud security firm Google acquired, runs Argon through its Scan for Good initiative to find exposures in public infrastructure. Google says Argon found a critical flaw in medical software used by hospitals worldwide, one that other frontier models missed. On Google’s internal confirmed-vulnerability dataset, Argon scored 85.8% versus 71.0% for Gemini 3.8 Flash Cyber.

The safety setup has four layers:

• Fairwind members and Google’s internal teams get a version without cyber guardrails. Everyone else will get a version that refuses cyber, chemical, biological, radiological, and nuclear attack requests.
• Probes monitor the model’s internal activations, not just its text output.
• A separate monitor can halt a task if Argon’s actions drift past what you asked for.
• Training included synthetic prompt injection attacks. On Gray Swan’s indirect prompt injection test, attacks succeeded against Argon 0.7% of the time at 15 attempts, against 8.5% for GPT-6 Astra.

Google has reason to be careful. During earlier autonomous cyber testing, a Gemini model accessed systems at three real companies instead of a fictional target because of weak sandboxing. Google says the model stopped itself and caused no harm, but the incident shaped how Argon ships.

The rest of the industry gates its top models the same way. Anthropic limits its Mythos models to Project Glasswing members, and OpenAI tiers GPT-6 Astra access through its Daybreak program.

How much will Gemini 4 Argon cost?

Google has set API pricing, even though the API isn’t open yet.

Google hasn’t said when the introductory period ends. Note the math on long outputs: a single 1 million token response costs $10 at the intro rate and $20 after it. Cached input is 95% cheaper than standard input, so reusing a large codebase or document set across calls saves real money.

Google also hasn’t published an official model ID. Developers should avoid hard-coding strings found on third-party sites.

How Gemini 4 Argon compares to Gemini 3

If you used Gemini 3, the jump isn’t about everyday chat. Gemini 3’s features focused on reasoning, multimodal understanding, and app integration for regular users. Argon targets long, multi-step professional work: code migration, financial analysis, legal review, and security research. It’s a specialist tool first and a consumer model later.

Final Thoughts

The 1 million token output gets the headlines, but Argon’s low hallucination rate may matter more to the people who’ll pay for it. It’s not the best coder on every chart, and most of you can’t touch it today. If your work involves long documents, financial analysis, or security audits, though, Argon is built to finish the whole job instead of the first chunk. Watch for the API and Ultra rollout. Would you trust an AI that refuses more often but makes things up less?