Gemini 4 Argon: Where to Get It and 5 Things to Test First
Google's new model came top on the Vals Index. Here's where to get it, what it costs, and five real jobs to try it on.
Back to all free guidesSo in the reel I went through Gemini 4 Argon, Google's new model that's built for the big jobs. This is the full update: where to get it, what it costs, and the five things I'd test it on first.
What it is, in a minute
- Google released it on 13 October 2026. They first showed it on 30 September, then only let cyber defenders use it for nearly two weeks.
- It keeps working way longer. Its output limit went from 64K tokens to a million, so it can think through a massive problem in one go.
- It came top of the Vals Index. That's the independent leaderboard that scores models on real work in finance, coding, legal and tax. Argon is first on 68.90%, just ahead of Claude Sonnet 5.5 on 67.04%, and it's also first on Vals' finance agent test.
- Google's demos are pretty wild. It turned pictures of a real desk into a racing game you can actually play, and built a whole pagoda temple in Blender, with agents checking each other's work.
You can play the desk racer yourself in AI Studio: https://aistudio.google.com/apps/bundled/desk-racer
Where to get it
- The Gemini app, if you're on Google AI Pro or Ultra. Pick Argon in the model picker.
- Google AI Studio, on a paid Gemini API key. The model's called
gemini-4-argon. - Google Antigravity, if you're on Ultra.
What it costs on the API
- $2 per million tokens in and $10 out, through 30 November 2026.
- $4 in and $20 out from 1 December 2026.
- It's on the paid tier only, so you'll need billing switched on for your key.
For comparison, GPT-6 Astra and Claude Fable 5.1 are both $10 in and $50 out, so until December Argon is a fifth of their price per token.
Five things to test it on first
Run each one on Argon and on whatever you use now, then compare the answers side by side.
1. A long video. Upload a recorded meeting, a webinar or a long screen recording. Long video is one of its strongest scores.
Watch this recording and give me chapters with timestamps, a two-line summary of each chapter, and every action item with who owns it.2. A messy folder into a client proposal. This is one of Google's own demos in the Gemini app.
Here are my notes, emails and files for this project. Turn them into a client proposal with the problem, what we'll do, the timeline and the price. Keep it to two pages and list anything you had to guess.3. What the numbers are really saying. Paste in a long report or a spreadsheet export. It came first on Vals' finance agent test, so this is the fairest place to look for a difference.
Here's our export for the last quarter. Tell me what's really going on, the three numbers I should worry about, and what you'd do next. Show the working for every figure.4. The waffle test. Google says it hedges less and skips the clichés. Ask it something you asked another AI last week, and count the paragraphs of "it depends".
5. Hand it a task. In the Gemini app, switch from Chat to Task and give it a job to get on with while you're away from your desk.
Research the five best-rated bookkeeping tools for a small service business. Compare the price, bank feeds, invoicing and how easy it is to switch over. Give me a table and your pick.
Business type:
Team size:Long jobs in the API
A normal API call stops well before a million tokens. Anything over 64K tokens of output needs background mode, where it works in 64K hops for up to 24 hours, and you pay for every hop. Google's guide is here: https://aistudio.google.com/docs/long-decode
Check what a job really costs
It's cheaper per token, but it uses a lot of tokens. Artificial Analysis measured about 62,000 output tokens a task against 27,000 for GPT-6 Astra. So run the same job on two models and compare the bill, not just the price per token.
To be honest, that's exactly what I'm doing next: a couple of real client jobs on Argon next to Claude, and I'll share what comes back.
Google's launch post: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/