What Is Gemini 4 Argon? Google's First Gemini 4 Model for Reasoning, Coding, and Cyber Defense

Gemini 4 Argon is a large language model Google announced on September 30, 2026, and the first model in the Gemini 4 generation. Google's developer docs call it the company's "flagship reasoning model," built for complex software engineering, hard math and science problems, and AI agents that carry out many steps on their own.

Its rollout was unusual. Google first released it only to cybersecurity partners, then opened it to developers and paying Gemini app subscribers on October 13. Google's reasoning: the model is strong enough at security work that defenders should get it first.

For everyday users, think of it as Google's most capable Gemini model right now, and the one best suited to long, difficult work.

If the Gemini Flash models are quick-thinking assistants, Argon is more like a senior engineer who can stay focused for hours and see a big project through.

The Short Version

Gemini 4 Argon is Google's new flagship reasoning model, with a 1-million-token context window and up to about 1 million tokens of output. It targets coding, long-running agents, and cyber defense. API prices are at a lower rate for now and will double on December 1, 2026.

What Problem Does It Solve?

AI models are increasingly used for jobs that run a long time, such as migrating an entire codebase, finding and patching software vulnerabilities, or analyzing long videos and stacks of documents.

Those jobs are hard in two ways:

1. The model has to hold a lot in mind: It needs to see large amounts of code or text at once.
2. The model has to produce a lot at once: Big rewrites or reports often exceed typical output limits.

Argon is designed to stretch both ends: it can read a lot, and it can write a lot in one go.

Key Features

1. Huge Context and Long Output

The context window is how much content a model can "see" at once. Argon's limit is 1,048,576 tokens, which is roughly a few hundred thousand words of English.

Output length is where it stands out. A standard request can return up to 65,536 tokens. With what Google calls long decode, output can reach about 1 million tokens. Note that input and output share one limit; you don't get 1 million of each.

2. Adjustable Thinking

Argon supports thinking at low, medium, and high levels. Deeper thinking usually produces more careful answers but takes more time and costs more. Low is plenty for simple questions.

3. Built for Cyber Defense

Google says Argon was trained specifically for defensive security work and can "autonomously find, validate, and patch critical software vulnerabilities." That's why it first went only to governments and trusted partners through Google's Fairwind Program.

Google also says Argon is better at resisting prompt injection, where someone hides malicious instructions in a web page or document to trick an AI into doing something the user never asked for.

4. Multimodal Input and Many Tools

Argon reads text, images, video, audio, and PDFs, but it only outputs text. It can't generate images or speech.

In the API, it supports function calling, grounding with Google Search and Google Maps, code execution, file search, and structured outputs. Computer use is available in preview.

Specs and Pricing

Standard API pricing (USD per 1M tokens):

Batch and Flex run at about half the standard rate, while Priority costs more. Only implicit (automatic) caching is supported; you can't set up caching manually. Prices may change, so check Google's pricing page before you commit.

How Does It Compare?

On performance, Google says Argon beats GPT-6 Astra and Anthropic's models on several benchmarks. Engadget, citing the independent benchmarking firm Artificial Analysis, reports that Argon matches GPT-6 Astra on its overall intelligence index at about 60% of the cost per task at current prices.

Other third-party roundups, however, note that Argon still trails Claude Opus 5.5 on Artificial Analysis's independent index. Benchmarks use different methods, so don't read any single ranking as proof of which model is "best."

What Should Non-Engineers Know?

• You can use it in the Gemini app: Google AI Pro and Ultra subscribers can pick Argon directly, no coding required.
• A different answering style: Google says Argon answers more directly, with fewer hedges and less repetition.
• You can hand it tasks: A new tasks mode in the Gemini app lets Gemini keep working while you're away, and it connects to more than 40 apps, including Instacart, OpenTable, Adobe, and Shopify.
• Antigravity on Mac: The macOS Gemini app includes Google Antigravity's coding interface, so you can switch between chatting and coding.

Who Is It For?

• Developers working on large codebases, multi-file refactors, or migrations.
• Researchers and knowledge workers analyzing long videos, long documents, or lots of charts.
• Teams building AI agents that run for long stretches on their own.
• Security teams responsible for finding and fixing vulnerabilities.

Who Might Not Need It?

• People who only need quick answers or short rewrites: the Gemini Flash models are faster and cheaper.
• Anyone hoping to use it for free: there's no free API tier, and the Gemini app limits it to paid plans.
• Apps that need image generation, speech, or real-time voice: Argon doesn't support the Live API or image and audio generation.

How to Get Started

Everyday users: Subscribe to Google AI Pro or Ultra, then choose Gemini 4 Argon from the model menu in the Gemini app.

Developers: Get an API key from Google AI Studio and set the model name to gemini-4-argon:

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-argon:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Summarize this project architecture in three points"}]}]}'

Enterprise customers can also access it through Model Garden on Google Cloud.

Our Take

What stands out about Argon isn't any single benchmark score. It's Google's rollout strategy: give it to security defenders first, then open it up gradually. That says a lot about how capable frontier models have become. Vendors now have to think about who gets access first.

Pricing matters too. At $2 per million input tokens and $10 per million output tokens, Argon is far cheaper than GPT-6 Astra right now, but those rates double on December 1. If you're evaluating it, this is a good window for testing, but budget with the post-increase prices so your costs don't surprise you after launch.

Finally, 1 million tokens of output sounds impressive, but generating that much at once is expensive and slow. In most cases, breaking a big job into smaller steps is still more reliable and much easier to review.

Sources

• Google blog: Gemini 4 Argon is now available in the Gemini app: https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/
• Gemini API model docs: https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
• Gemini API pricing: https://ai.google.dev/gemini-api/docs/pricing
• TechCrunch: https://techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet/
• Engadget: https://www.engadget.com/2274263/google-gemini-4-model-argon/
• Accessed: 2026-10-21