Gemini 4 Argon Tutorial: Setup, Pricing, and Benchmarks

Google put Gemini 4 Argon in people's hands on October 13, 2026. A week earlier, on September 30, it had only been open to a small group of cyber defenders in the Fairwind Program. If you pay for the Gemini API, subscribe to Google AI Pro or Ultra, or use Gemini for business, you can try it today.

This is a practical walkthrough: what Argon actually is, where you can use it, what a request costs, how to call the model from Python, and how Google's own October 2026 benchmark table compares it with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5.

What Gemini 4 Argon is

Argon is the first model in the Gemini 4 family. Google describes it as a large sparse mixture-of-experts transformer with native support for text, images, audio, and video on the way in, and text on the way out. It is not a fine-tune of an older Gemini. The public model id is gemini-4-argon.

The context window is one shared budget of 1,048,576 tokens. Input and output both draw from that pool, and output includes thinking tokens and function calls, not just the words you see. A normal request stops at 65,536 output tokens (32,768 if you call the non-streaming HTTP API directly). Longer answers use a continuation token. The knowledge cutoff on the model card is March 2026.

It is built for work that takes many steps: code migrations, legal and finance documents, science and math, and long agent runs. It will not generate images or audio, and it is not on the Live API. Caching is implicit only. If a shared prompt prefix is at least 6,144 tokens, Google discounts those cached input tokens by 95 percent. Put the static part of your prompt first.

Who can use it right now

• Developers. Paid Gemini API in Google AI Studio and Android Studio. Google AI Ultra subscribers can also open it in Google Antigravity.
• Companies. Gemini for business, the API, and Antigravity. On Google Cloud the model id is the same, and the supported locations are global, us, and eu.
• Gemini app and Search. Google AI Pro and Ultra subscribers get Argon in the Gemini app and in AI Mode in Search.
• Researchers. Google is taking applications for 10,000 scientists to receive Antigravity's highest usage tier free for six months, subject to review and capacity.

A separate model id, gemini-4-argon-cyber-permissive, exists for organizations already accepted into Fairwind. Everyday apps should call gemini-4-argon.

Gemini 4 Argon pricing

Public pricing has two phases. Through November 30, 2026 the introductory rate is $2 per million input tokens and $10 per million output tokens. After that date it becomes $4 per million input and $20 per million output. Cached input stays at 95 percent off the input price. Thinking tokens are billed. Google says this schedule applies to the public Argon model.

The Gemini app does not bill by the token if you are on Pro or Ultra. API usage does. If you are prototyping, start on LOW or MEDIUM thinking so a long reasoning trace does not surprise you on the invoice.

How to call Gemini 4 Argon

You need a paid Gemini API key from Google AI Studio, or Application Default Credentials on a Google Cloud project. Install a current SDK. Older clients cannot follow Argon's long completions.

• Python: google-genai 2.30.1 or newer
• JavaScript: @google/genai 2.29.1 or newer
• Go 1.74.0, Java 1.77.0, .NET 1.26.0, Kotlin 1.6.0

pip install --upgrade "google-genai>=2.30.1"

A first Python request

Set GEMINI_API_KEY in your environment. Do not paste the key into source code. MEDIUM is the default thinking level on the public model, so you can omit it, but setting it makes the choice obvious when you come back to the file later.

import os
from google import genai
from google.genai import types

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

response = client.models.generate_content(
    model="gemini-4-argon",
    contents="Review a checkout flow that writes orders to one primary database at 15,000 writes per second. Name the first bottleneck.",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_level="MEDIUM"),
    ),
)
print(response.text)

On Google Cloud, the same call uses an enterprise client instead of an API key. Google's developer guide uses location="global":

client = genai.Client(
    enterprise=True,
    project="YOUR_PROJECT_ID",
    location="global",
)

Pick a thinking level on purpose

thinking_budget and the old MINIMAL level both return HTTP 400. The only valid values are LOW, MEDIUM, and HIGH.

Argon also ignores custom temperature, top_p, and top_k. Sending frequency_penalty, presence_penalty, or candidate_count fails the request. Control depth with thinking_level, and control shape with a response schema.

Ask for JSON

Structured output still works. This pattern is the one in Google's Argon guide, trimmed to a small schema:

from pydantic import BaseModel, Field
from google import genai
from google.genai import types

class Review(BaseModel):
    bottleneck: str = Field(description="The first thing that breaks")
    changes: list[str] = Field(description="Concrete changes, in order")

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])
response = client.models.generate_content(
    model="gemini-4-argon",
    contents="Review a synchronous order pipeline on one primary database at 15,000 writes per second.",
    config=types.GenerateContentConfig(
        response_mime_type="application/json",
        response_schema=Review,
        thinking_config=types.ThinkingConfig(thinking_level="HIGH"),
    ),
)
print(response.text)

The same call over HTTP

Replace the token with an AI Studio key, or with a Cloud access token if you are calling Vertex. Keep the generation config this small. Extra sampling fields are what usually break a migration from Gemini 2 or 3.

curl -X POST \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-argon:generateContent" \
  -d '{
    "contents": [{
      "role": "user",
      "parts": [{"text": "Give me a short architecture review of a single-database checkout."}]
    }],
    "generationConfig": {
      "thinkingConfig": {"thinkingLevel": "MEDIUM"}
    }
  }'

When a reply needs more than 65,536 tokens

Most coding and chat turns finish inside the normal cap. Use long decode for proofs, hard research, and the rare agent turn that thinks past 65,536 tokens. Set max_output_tokens above 65,536 on a streaming call (above 32,768 on a plain HTTP call), up to the remaining room in the 1,048,576 budget. Current Google Gen AI SDKs follow the continuation token for you. If you write your own loop, resend the original prompt plus the latest token. Do not paste the partial answer back in and ask the model to continue.

One catch from Google's guide: continuation on generateContent is on the v1beta1 API, not v1. The SDK already uses v1beta1 unless you override it. Long decode is also the wrong tool when you simply want a longer essay. It gives the model more room to think. It does not make it obey a word count.

Gemini 4 Argon benchmark table

These figures come from the Gemini 4 Argon model card, published October 13, 2026. They are Google's measurements against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. A dash means Google did not publish a score for that model. Higher is better on every row below. Methods are on DeepMind's evals page for this model, and independent labs will not match every number.

Read the table as a spread, not a single winner. Argon leads the Vals Index, which Google uses as a stand-in for finance, coding, legal, and tax work. It also leads AutomationBench, both finance and legal agent tests in this card, DeepSWE v1.1, Vibe Code Bench, LABBench 2, RiemannBench, both GraphWalks lengths, Chartography, and LVBench. The long-context gap gets wide once prompts pass 256k tokens: 84.2 percent versus 71.8 percent for GPT-6 Astra.

It does not lead everywhere. GPT-6 Astra is ahead on FrontierSWE v2, Terminal-Bench Science, and OSWorld-2.0. Claude Opus 5.5 is ahead on Terminal-bench 4.0 and PostTrainBench. CWE-bench v1 is a tie with GPT-6 Astra at 68 percent. Harvey's legal scores look small for every model because the task is hard. Argon is still the highest of the four, at 19.6 percent.

Inside Google, the launch post also cites a few product results that are not in the table: more than 35 percent more agentic code submissions in Antigravity, with rollbacks down by nearly a third, and an 87 percent score on the Embodied Reasoning benchmark. Treat those as Google's account of its own rollout.

What to use it for, and what to skip

Reach for Argon when the job is a pile of files, a long PDF, a proof, or an agent that has to keep a plan straight for many steps. The model card lists multi-turn investigations, engineering work, visual documents, scientific reasoning, and legal or financial writing as the intended use.

Skip it for cheap classification, live voice, image generation, and anything that has to answer in a few hundred milliseconds. LOW still spends thinking tokens. A smaller Gemini model will be the better default for a support chatbot that answers the same questions all day.

Argon still hallucinates, and the card warns about occasional slowness and timeouts. Its training cutoff is March 2026, so ask it to look things up with search grounding when the fact is newer than that. Check citations the same way you would with any other frontier model.

A short migration checklist

1. Change the model string to gemini-4-argon.
2. Upgrade the Google Gen AI SDK to the versions listed above.
3. Replace thinking_budget with thinking_level.
4. Delete temperature, top_p, top_k, penalties, and candidate count.
5. Remove prefilled assistant turns from old chat histories. Argon rejects them.
6. Drop explicit cache calls. Keep stable instructions at the top of the prompt so implicit caching can hit.
7. If you need outputs past 65,536 tokens, stream on v1beta1 and let the SDK handle continuation.

Common questions

Is Gemini 4 Argon free?

In the Gemini app and AI Mode, it is included for Google AI Pro and Ultra subscribers. The API is paid. The introductory API price runs through November 30, 2026.

What is the model name in code?

gemini-4-argon. That is the stable id in the Gemini API docs as of October 2026.

How long is the context window?

1,048,576 tokens, shared by input and output. A single ordinary response is capped at 65,536 output tokens unless you opt into long decode.

Can I still set temperature?

No. The backend ignores temperature, top_p, and top_k. Use thinking level and, when you need a fixed shape, a JSON schema.

When did Google launch it?

September 30, 2026 for a limited Fairwind group. October 13, 2026 for developers, businesses, and Pro and Ultra subscribers.

Sources: Google's October 13 launch post, the Gemini API model page, the Cloud developer guide for Gemini 4 Argon, and the DeepMind model card dated October 2026.