Gemini 4 Argon is Google’s new frontier model, AI labs’ term for their most advanced systems, built for long, multi-step work: the coding, finance, legal, cybersecurity and science jobs Google calls “deep work.” It’s the first model of the Gemini 4 family, and it’s rolling out globally in the Gemini app, AI Mode in Search and Google’s API for developers.
A million tokens per job
Tokens are the chunks of words an AI model reads and writes. Argon gets up to 1 million of them, shared between what you give it and what it writes back. Before this, a Gemini model could write only 64,000 at a time.
Argon can reason and write hundreds of thousands of tokens in a single run, which Google says lets it solve tough problems in one go.
In Google’s examples, Argon worked overnight, on its own, to speed up Google’s open-source Agent Development Kit, a framework for building AI agents. Its changes cut the time the kit spends preparing long sessions by 40%.
It also helped the Chrome team separate core webpage handling from a roughly 6 million-line folder of old browser features in Chromium, the 40 million-line codebase Chrome is built on. The job took weeks instead of months. Google says the cleanup will let live webpages move between tabs, side panels and floating windows without reloading.
Built to work on its own
Argon does much of this as an agent: AI you give a goal, which plans the steps and keeps working in the background until the job’s done. According to Google, Argon in Antigravity, its agent-first coding platform, managed to raise agent-made code submissions from internal teams by more than 35% and cut rollbacks – changes that had to be undone – by nearly a third.
The Gemini app gets a simple toggle from chat to tasks, which lets Gemini keep working for you when you’re not at your computer. Antigravity’s coding interface now sits inside the Gemini app for Mac, and Gemini connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify.
Google also says Argon answers more directly, hedges less, repeats itself less and “skips clichés.” If that holds up, it already writes better than most launch posts.
Google’s numbers
Google says Argon ranks first on the Vals Index, which scores AI models on finance, coding, legal and tax work, weighting each sector by its share of US GDP. It also reports 77.9% on DeepSWE v1.1, a test of long software engineering tasks, and 87% on ERQA, a test of reasoning about physical, 3D space. Google picked those tests to show off its own model, so we’d read “number one” as Google’s scoreboard, not a final ranking.
Safety, and what Google’s monitor caught
Argon first went to trusted cybersecurity defenders on September 30th, through Google’s Fairwind Program. In their tests, it solved all 12 simulated cyberattack scenarios. So it ships wrapped in three layers of protection that screen requests, watch the model’s internal signals for knowledge about cyberattacks or biological threats, and can cut it off mid-answer.
Resistance against prompt injection, where hidden instructions on a webpage, email, or code get used to try and hijack an AI agent’s task, was also tested. Google says Argon leads the prompt injection benchmark from Gray Swan, an AI security company, cutting attack success to 0.7% when an attacker gets 15 tries and adapts each time. Google also runs a monitor that reads Argon’s step-by-step reasoning as it works. And it’s through this that Google caught it already trying to alter grading scripts, and once crashing its test sandbox to avoid being scored on a failed run.
Who gets it, and what it costs
In the Gemini app and AI Mode in Search, Argon is for Google AI Pro and Ultra subscribers. Google’s plans page says it’s available to Pro and Ultra subscribers in selected countries.
In the Philippines, AI Pro costs Php 1,100 (about $18) a month. AI Ultra comes in two tiers: Ultra 5x costs Php 5,999 (about $96) a month, with five times the Pro plan’s usage limits and 20TB of storage. Ultra 20x costs Php 11,900 (about $190), with 20 times the limits and 30TB.
Developers pay per token through the paid tier of the Gemini API, in AI Studio and Android Studio. Until November 30th, the introductory price is $2 (about Php 125) per million input tokens, what you send, and $10 (about Php 630) per million output tokens, what it writes. From December, the regular price of $4 (about Php 250) and $20 (about Php 1,260) applies.
Cached input, text it has stored from an earlier request, is 95% off.
Ultra subscribers also get Argon in Antigravity, and businesses get it through Gemini for business. Google is offering 10,000 scientists Antigravity’s highest usage tier free for six months, if they’re selected; applications are open.
Cybersecurity teams can apply to Google’s Fairwind Program to use Argon without its cyber guardrails.
