Gemini 4 Argon: Google's New Frontier AI, Explained
By Chatday Editorial Team ·
Inside Google, a team had a job nobody wants: rewrite a huge chunk of old C and C++ code, more than 800,000 lines of it, into a newer, safer programming language called Rust. That kind of project normally eats months of careful, boring work.
Google says it handed a large part of that job to its newest AI model. The model is called Gemini 4 Argon, and Google announced it on September 30, 2026.
A model built for the jobs that take all day
Most of what people ask an AI takes seconds: a quick answer, an email draft, a recipe. Argon is pointed at the other kind of work. Google calls it a frontier model (its most advanced tier) for long, multi-step tasks in three areas: real-world software engineering, office knowledge work such as legal and finance, and cybersecurity defense.
The code rewrite isn’t the only example Google gave. It says Argon also helped free up more than 300 TiB of memory across its data centers, and rewrote a video decoder in Rust that runs 2.7 times faster than the earlier Rust version. These are Google’s own internal results, so treat them as a showcase rather than a promise. But they show what “long, real-world work” means here: projects, not questions.
There’s a naming twist, too. Earlier Gemini models came in Pro and Flash versions. Argon is the first Gemini to carry a codename instead, and the first in a new Gemini 4 family. Google hasn’t explained the name.
You can check its full specs on the Gemini 4 Argon model page.
How Argon can write a whole book in one go
The most unusual number in the launch is about output, meaning how much the model can write back in a single answer. Google says Argon can produce up to 1 million tokens (a token is a small chunk of a word) in one response, up from 64,000 on earlier models. That’s room for something the length of several novels.
There’s a catch worth knowing. The question you send, the model’s private thinking and its answer all share the same window of about 1 million tokens. And by default, Google’s developer tools stop a single answer much sooner, at around 65,000 tokens, unless an app switches on a special long-output mode. So for normal chatting you won’t notice a difference. It matters when an app asks Argon to produce something truly huge, like a full codebase or a long report.
| What you’d ask | What that means for Argon |
|---|---|
| A quick question or a short draft | Works like any strong chat model |
| A long document to read and analyze | Reads about 1M tokens at once |
| A very long piece of writing or code | Can keep writing far longer than most models, if the app allows it |
| A screenshot, chart or PDF | Understands images and files alongside text |
What the test scores actually say
Google published plenty of benchmark wins. In plain terms: on a test of long, real-world software tasks, Argon came out on top, slightly ahead of Anthropic’s Claude Opus 5.5. It also led tests on office automation and on legal work.
Outside reviewers are more careful. Artificial Analysis put Argon a hair behind the leader on its overall intelligence ranking, essentially a tie. And The Decoder reported a mixed picture on accuracy: Argon makes things up less often than some rivals, but it also answered fewer factual questions correctly in their test. So it’s a strong model, not a magic one, and it trails rivals on a couple of coding tests, including one focused on command-line work.
The security side: an AI that hunts bugs
Cybersecurity is the part of the launch Google talked about most carefully. Argon can find, confirm and patch serious software flaws on its own. Google says that, working with the security firm Wiz, it found a critical flaw in healthcare software that earlier models had missed.
Because that skill could be misused, Google is cautious. At launch it gave the full, unrestricted version only to vetted defenders through a program called Fairwind. Its safety report says it can’t rule out that Argon crosses Google’s own danger thresholds for cyber and biology, so extra safeguards are switched on. For everyday writing, research or coding, none of this gets in your way.
What it costs, and who can use it
Google is charging developers an introductory $2 per million tokens sent in and $10 per million out, until November 30, 2026. After that, the price is set to rise to $4 and $20. Inside Google’s own consumer plans, Argon isn’t part of the free tier or the Google AI Pro plan.
You don’t need any of that to try it. Gemini 4 Argon is available in Chatday, so you can open it in the same window where you already chat with GPT, Claude or Grok, and switch between them mid-conversation. If you’re wondering how it stacks up against a faster, lighter model for everyday jobs, our guide to fast vs smart AI models walks through the trade-off.
It’s also a reminder of how big Gemini has become: we covered the moment Gemini passed a billion users.
The best way to judge Argon is to give it the kind of job it was built for. Pick something long and annoying (a messy document, a big piece of code, a report you’ve been dreading) and see how far it gets before you have to step in.