How to Use Gemini 4 Argon (and How to Get Access)

Google announced Gemini 4 Argon on September 30, 2026, but most people can't use it yet. Here's who gets access first, the two routes in (the Gemini app and the Gemini API), what to set up today so you can switch on day one, and how to prompt a model built for long, heavy jobs.

Who gets Gemini 4 Argon first

Google is rolling Argon out in stages and says it will widen access gradually while it gathers feedback. Here's the order as announced:

Google's announcement doesn't describe a way to apply for Fairwind, so for most people the practical routes are the Gemini app with AI Ultra or a paid Gemini API account. For the latest on timing, see our Gemini 4 Argon release date page.

Route 1: the Gemini app with Google AI Ultra

1. Have Google AI Ultra. Google names AI Ultra subscribers among the first people to get Argon. It hasn't mentioned lower plans.
2. Open the Gemini app. Sign in on the web or on your phone with the account that holds the subscription.
3. Check the model menu. Google hasn't said exactly where Argon will appear. Since access widens gradually, a colleague may see it before you do.
4. Start with a big task. Argon is built for long, multi-step work. A one-line question won't show you much of what it can do.

Route 2: Google AI Studio and the Gemini API

1. Get an API key. Sign in to Google AI Studio. It creates a project and an API key for new users automatically, and you can copy yours from the API keys page.
2. Be a paid customer. Argon's first API users are paid API customers, so plan on having billing set up for your project.
3. Watch the models page. Argon isn't listed yet. The newest public IDs there include gemini-3.8-flash.
4. Swap the ID and test. When the Argon ID appears, change one config value and run your test set. Our Gemini 4 Argon API page covers the details.

What about Vertex AI?

What to do today so you're ready on day one

1. Get a Gemini API key. From Google AI Studio, as above.
2. Build against the current model. Keep the model ID in config, not scattered through your code, so switching to Argon is a one-line change.
3. Write a small test set. Ten to twenty real tasks with answers you know are good. On launch day you can compare Argon with your current model in an afternoon instead of trusting headline scores (Google's own numbers are on our benchmarks page).
4. Budget. Google's intro pricing is $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% off, rising later to $4 and $20. See pricing.

Here's the minimal Python setup with Google's Gen AI SDK. Install it with pip install -U google-genai and set GEMINI_API_KEY in your environment:

import os
from google import genai

# Swap this to the Gemini 4 Argon ID when Google publishes it.
MODEL = os.environ.get("GEMINI_MODEL", "gemini-3.8-flash")

client = genai.Client()  # reads GEMINI_API_KEY
interaction = client.interactions.create(
    model=MODEL,
    input="Summarize this changelog for customers: ...",
)
print(interaction.output_text)

Already using the OpenAI SDK? Gemini has an OpenAI-compatible endpoint, so you only change the key, the base URL and the model. Google notes this compatibility layer is still in beta.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["GEMINI_API_KEY"],
    base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
)
response = client.chat.completions.create(
    model=os.environ.get("GEMINI_MODEL", "gemini-3.8-flash"),
    messages=[{"role": "user", "content": "Explain to me how AI works"}],
)
print(response.choices[0].message.content)

Want Argon inside your coding assistant? Our guide to connecting Gemini to Claude sets it up as a tool Claude can call.

How to prompt Gemini 4 Argon

Google built Argon for long-horizon coding and enterprise knowledge work, and says it can write up to 1M tokens in a single answer, up from 64K. That changes how you should ask.

Ask for the whole thing

With that much output room, you no longer need to break work into small pieces. Ask for every changed file in full, the complete report or the entire test suite in one go.

Give it the full context

Include everything relevant: the spec, the code, the data and an example of output you like. Reusing the same large context also gets cheaper, since Google prices cached input at 95% off.

Ask for a plan first

For big jobs, have Argon write a numbered plan, check it yourself, then say go. Fixing a plan is much cheaper than fixing a million tokens of output.

Define "done"

Spell out acceptance criteria, format and what to skip, then ask Argon to check its work against that list at the end.

Cap the output when you don't need it all

At the intro price, a maximum-length answer costs about $10 in output tokens alone. Set a maximum output length for everyday calls.

Here's a plan-first prompt you can adapt:

Goal: Move our billing module from raw REST calls to the new SDK.
Context: [paste the module, the SDK docs and our style guide]

Step 1: Write a numbered plan. List every file you will change and why.
Then stop and wait for my go-ahead.

Step 2 (after I say go): Output every changed file in full, one after
another, each headed by its path.

Done means: existing tests still pass, no public function signatures
change, and you end with a list of anything you were unsure about.

Images, video and music: use what's ready now

Google hasn't said Argon makes images, audio or video, so for creative work you don't need to wait. Lumeta AI has these tools today (and our video guide shows how to plan a clip with a language model and make it with a video model):

Frequently asked questions

Sources: Google (Gemini 4 Argon announcement), 9to5Google, Gemini API quickstart, Gemini API models, Gemini API OpenAI compatibility, Gemini Interactions API.