Gemini 4 Argon: What It Is, What It Costs and What People Are Doing With It

What Gemini 4 Argon is

Gemini 4 Argon is the first model in Google DeepMind's Gemini 4 line. Google's launch post positions it for long-horizon software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. Its headline spec is a 1 million token output limit, up from 64K on the previous generation according to 9to5Google. Progressive Robot says this relies on a feature that resumes a response across several requests, and that its own testing saw roughly 262,000 tokens in a single request.

Who makes it and when it arrived

Google DeepMind announced Argon on 30 September 2026. A stealth checkpoint had appeared on Arena a few days earlier under a different name, as Pasquale Pillitteri reported, though the Argon codename was only a leak at that point.

Access and pricing

Access is gated. Google says Argon is rolling out first to trusted cyber defenders in its Fairwind Program, with wider availability to developers, enterprises and consumers to follow, starting with paid API customers and Google AI Ultra subscribers. Google lists introductory pricing of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. Its post gives later rates of $4 and $20. No firm date for general API access has been given; Progressive Robot says only "as soon as possible". NeuralTrust noted Argon was absent from OpenRouter, Vertex AI and GitHub Copilot at launch.

What people build and do with it

Because few people can use Argon, public work is mostly demonstrations, partner stories and evaluations rather than hands-on projects.

Cyber defense. This is the flagship use. Google cites Wiz using Argon to find a critical healthcare software flaw that exposed personal data. Wiz's own Scan for Good page describes a wider effort to scan public-service and infrastructure systems, but it names an earlier Gemini cyber model rather than Argon, so treat it as context.

Large code work. Google and press coverage point to big code migrations and infrastructure optimization inside Google. NeuralTrust points out that single answers this long are difficult for humans to review.

Knowledge work. Reviews such as eesel's find Argon strongest on legal and financial documents and multi-step business automation.

Benchmarks and analysis. Most community activity is evaluation. Artificial Analysis scored Argon 53 on its Intelligence Index, level with GPT-6 Astra, and reported a 15% hallucination rate. The Decoder adds that Argon uses far more output tokens per task and that Claude models still score higher on that index.

Video explainers. Creators such as Prompt Engineering, Sam Witteveen and AI with Surya have published overviews.

Limits and caveats

• Coding is mixed. Google reports 77.9% on DeepSWE v1.1, but a DEV Community breakdown reports an 8th-place Code Arena WebDev rank and weaker from-scratch coding results.
• Real cost may differ from the sticker price. Argon's higher token use narrows the discount advantage, per eesel and The Decoder.
• Security claims are thin. NeuralTrust notes Google said it leads a prompt-injection benchmark without publishing a score.
• Internal doubts. Aroged reports, via unnamed sources, that some Google staff question real-world coding performance. That is unconfirmed.
• Little independent validation. With access restricted, most numbers come from Google or a few early evaluators. Outlets also differ on early access: 9to5Google mentions Ultra subscribers, while Google's own post says Fairwind first.

Frequently asked questions

Can I use Gemini 4 Argon today?

Only if you are an approved Fairwind Program cyber defender. Google says wider access will follow, starting with paid API customers and Google AI Ultra subscribers, without a date.

How much does it cost?

Introductory pricing is $2 per million input tokens and $10 per million output tokens. Google's post lists $4 and $20 as the later rates.

What is the output limit?

One million tokens, according to Google. Independent testing reportedly saw about 262,000 tokens per single request.

Is it better than GPT-6 Astra or Claude Opus 5.5?

It depends on the task. Artificial Analysis rates it level with Astra overall, while Claude models score higher on that index; Argon leads on several knowledge-work and long-context tests.