Gemini 4 Argon: How to Use the Model and API

👁 124 reads

Gemini 4 Argon has been available since October 13, 2026, but access varies by route. Google lists four options: the Gemini app with Pro and Ultra plans, the paid Gemini API in AI Studio, Antigravity for Ultra subscribers, and OpenRouter, where the model also appeared. Below, I explain each option and share two Python scripts. I tested the OpenRouter version from Poland on the evening of October 13, but could not run the Gemini API version; I explain why. For a full overview of the model, see the Gemini 4 guide.

Which access option should you choose?

Option 1: Gemini app

This is the simplest route, but it requires a paid subscription. The free plan includes only 3.5 Flash-Lite, while Google AI Plus does not include Argon. You need Google AI Pro or Ultra.

1. Go to gemini.google.com or open the Gemini app on your phone.
2. If you do not have a plan, subscribe to Google AI Pro or Ultra. Prices vary by market.
3. Click the model selector above the input field. Google says Pro and Ultra include 3.5 Flash-Lite, 3.8 Flash, 3.1 Pro and 4 Argon.
4. Select 4 Argon and start chatting.

Availability in practice may differ: Argon was not yet on my Pro account in Poland on the evening of October 13. The model selector showed only 3.5 Flash-Lite, 3.8 Flash and 3.1 Pro. Google says the model is “starting to roll out,” so access is being phased in. See Gemini 4 Argon missing from the app for checks to make before troubleshooting.

Argon is also due to come to AI Mode in Search for Pro and Ultra subscribers. Google has not provided a list of countries.

Option 2: Gemini API in Google AI Studio

Google’s API model is named gemini-4-argon and is listed as stable in the documentation. It is available only on the paid API tier. Here’s how to get started:

1. Go to aistudio.google.com and sign in with your Google account. For more on the tool, see the Google AI Studio guide.
2. Click Get API key and create a key in a new or existing Google Cloud project.
3. Enable billing for the project. Without billing, you are on the free tier, and Argon is available only on the paid tier.
4. Save the key in the GEMINI_API_KEY environment variable.
5. Use the model ID gemini-4-argon in your code; see the script below.

Prices per 1 million tokens are USD 2 for input and USD 10 for output, with cached input at USD 0.10. These introductory prices apply through November 30, 2026. From December 1, prices rise to USD 4 / 20, with cached input at USD 0.20. The shared context window allows up to 1 million input and output tokens combined.

One caveat: according to 9to5Google, testing Argon in the AI Studio interface (the browser chat window) requires an Ultra plan. The paid API is available to anyone with a paid Gemini API account.

Option 3: OpenRouter

OpenRouter is a third-party service that provides access to multiple models through one API key. Argon appeared there on October 13 as google/gemini-4-argon, at the same price as Google: USD 2 / 10 per 1 million tokens, with a context window of 1,048,576 tokens. I used this route for all my tests.

1. Create an account at openrouter.ai.
2. Add credit by card in the Credits tab. A few dollars should last a while: my 12 test prompts cost USD 0.13 in total.
3. To use it without coding, open Chat, search for “Gemini 4 Argon” in the model selector and start chatting.
4. To use it in code, create a key in the Keys tab and use the script below.

Option 4: Antigravity

Antigravity is Google’s coding agent, available as a desktop app and CLI. Argon is available in it only to Google AI Ultra subscribers. Pro includes Antigravity access, but with other models. See Getting started with Antigravity for installation and first steps. Researchers can apply for six months of the highest Antigravity tier at no cost; Google accepts applications on an ongoing basis.

Ready-to-use script: Argon through OpenRouter

Download both scripts (ZIP, 2 KB): argon_openrouter.py (tested October 13, 2026) and argon_gemini_api.py (written to match the documentation, not run).

The script sends one question, lets you select a reasoning level, and prints the response, token counts, cost and runtime. I tested it on Windows 11 with Python 3.13 on October 13, 2026, in a clean folder containing only the script.

1. Install Python 3.10 or later and the library: pip install requests.
2. Set your API key. In PowerShell: $env:OPENROUTER_API_KEY="your-key".
3. Save the code as argon_openrouter.py.
4. Run it: python argon_openrouter.py "Your question" low.

# Gemini 4 Argon through OpenRouter (tested October 13, 2026)
# Requires: Python 3.10+, pip install requests
# Key: OPENROUTER_API_KEY environment variable
# Usage: python argon_openrouter.py "question" [low|medium|high]
import os
import sys
import time
import requests

sys.stdout.reconfigure(encoding="utf-8")

MODEL = "google/gemini-4-argon"
URL = "https://openrouter.ai/api/v1/chat/completions"

# Translated version of the original Polish test prompt:
# "What is 17% of 2350? Answer in one sentence."
question = sys.argv[1] if len(sys.argv) > 1 else (
    "What is 17% of 2350? Answer in one sentence.")
level = sys.argv[2] if len(sys.argv) > 2 else "low"

key = os.environ.get("OPENROUTER_API_KEY")
if not key:
    sys.exit("OPENROUTER_API_KEY environment variable is missing")

start = time.time()
r = requests.post(
    URL,
    headers={"Authorization": "Bearer " + key},
    json={
        "model": MODEL,
        "reasoning": {"effort": level},
        "usage": {"include": True},
        "messages": [
            {"role": "user", "content": question},
        ],
    },
    timeout=300,
)
r.raise_for_status()
d = r.json()
elapsed = time.time() - start

print(d["choices"][0]["message"]["content"])

u = d.get("usage", {})
details = u.get("completion_tokens_details") or {}
print()
print("model:", d.get("model"))
print("reasoning level:", level)
print("input tokens:", u.get("prompt_tokens"))
print("output tokens:", u.get("completion_tokens"))
print("including reasoning:", details.get("reasoning_tokens"))
print("cost (USD):", u.get("cost"))
print(f"time: {elapsed:.1f} s")

Here’s what I got in my test:

• With no arguments (low level): “17% of 2350 is exactly 399.50.” This is a translation of the original Polish test response, not a separate English-language test. The run used 51 input tokens and 25 output tokens, cost USD 0.000352 and took 2.2 seconds.
• Medium-level question about the difference between two types of Polish civil-law contracts: three factual sentences, 154 output tokens, USD 0.001648 and 3.0 seconds. The original test was in Polish; this result does not establish performance on English prompts.

OpenRouter reported zero reasoning tokens in both cases, including at medium. The same was true in all my low-level tests, which is why the responses were short and inexpensive. The sys.stdout.reconfigure line is there to prevent Windows consoles from failing on Unicode characters.

Using the official Gemini API

This script does the same thing through Google’s google-genai SDK. To be clear, I did not run it because I do not have a paid-tier Gemini API key. I wrote it to match the SDK documentation, and it estimates cost from Google’s pricing because the SDK returns token counts, not a charge amount. If it does not work, first check that billing is enabled for the project.

# Gemini 4 Argon through the official Gemini API
# (google-genai SDK). NOT RUN - no paid-tier key.
# Written according to the SDK documentation.
# Requires: Python 3.10+, pip install google-genai
# Key: GEMINI_API_KEY environment variable (from AI Studio)
# Usage: python argon_gemini_api.py "question" [low|high]
import sys
from google import genai
from google.genai import types

sys.stdout.reconfigure(encoding="utf-8")

MODEL = "gemini-4-argon"
# Introductory prices through November 30, 2026, in USD per 1 million tokens
INPUT_PRICE, OUTPUT_PRICE = 2.0, 10.0

# Translated version of the original Polish test prompt:
# "What is 17% of 2350? Answer in one sentence."
question = sys.argv[1] if len(sys.argv) > 1 else (
    "What is 17% of 2350? Answer in one sentence.")
level = sys.argv[2] if len(sys.argv) > 2 else "low"

# Client() reads GEMINI_API_KEY from the environment
client = genai.Client()
response = client.models.generate_content(
    model=MODEL,
    contents=question,
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_level=level,
        ),
    ),
)

print(response.text)

metadata = response.usage_metadata
input_tokens = metadata.prompt_token_count or 0
output_tokens = metadata.candidates_token_count or 0
thinking_tokens = metadata.thoughts_token_count or 0
# Thinking tokens are charged at the output rate
cost = (input_tokens * INPUT_PRICE + (output_tokens + thinking_tokens) * OUTPUT_PRICE) / 1e6
print()
print("input tokens:", input_tokens)
print("output tokens:", output_tokens, "| thinking:", thinking_tokens)
print(f"estimated cost: {cost:.6f} USD")

From December 1, 2026, change the prices in the script to 4.0, 20.0.

Frequently asked questions

Is Gemini 4 Argon free?

No. The free plan includes 3.5 Flash-Lite. Argon requires Google AI Pro or Ultra, or the paid API tier.

Do you need a VPN?

I did not need one for OpenRouter when using a Polish account. Google does not exclude Poland from the app or API, and Argon’s absence from my account appeared to be part of the gradual rollout, not a regional block. Availability can vary by country, so check with the provider.

Which reasoning level should you choose?

Start with low. At that level, Argon scored 99.5% on my 23-task comparison, at a cost of USD 0.22.

Kinetyra