Gemini 4 Argon Goes to Cyber Defenders First, With Broad Access Undated

Google released Gemini 4 Argon on Sept. 30, but only to a limited group: trusted cyber defenders enrolled in its Fairwind Program. The company said broader access will follow "as soon as possible", starting with paid API customers and Google AI Ultra subscribers, and gave no date.

The announcement came in a Google blog post written by Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect. TechCrunch reported the same day that the rollout is limited to cyber partners.

What Argon costs and how much it can read

Google listed an introductory price of $2 per million input tokens and $10 per million output tokens. After the introductory period the price becomes $4 and $20. Google did not say when that period ends.

Cached inputs, where a repeated prompt prefix is stored and reused, carry a 95 percent discount off the input price. The context window is 1 million tokens. The output limit is also 1 million tokens, up from 64,000 on the previous Gemini limit, according to the post.

Item	Introductory	After introduction
Input, per million tokens	$2	$4
Output, per million tokens	$10	$20
Context window	1 million tokens	1 million tokens

The Vals AI model page already lists the $4 and $20 rates as its base price.

Who ran the benchmarks

Most of the scores in Google's post come from Google itself. It reported 77.9 percent on DeepSWE v1.1, a software-engineering test, and 91.7 percent on LVBench, a long-video test. Both are vendor-supplied figures.

Third parties are named for others. Zapier's AutomationBench put Argon first at 51.3 percent, according to Google. On CWE-bench v1, which measures finding software weaknesses, Google said Argon scored 68 percent and tied for first.

The independent evaluator Vals.ai lists Argon first of 43 models on its Vals Index at 68.90 percent, with a margin of plus or minus 0.97. It ranks Argon first of 74 on Finance Agent v2 at 65.40 percent and second of 44 on CyberBench at 77.86 percent.

Vals also shows where Argon is weaker. It ranks the model second of 108 on Vibe Code Bench at 91.91 percent and third of 149 on LegalBench at 88.30 percent. On medical transcription it scores 87.43 percent and ranks 15th.

1. Gemini 4 Argon77.9 %
2. GPT-6.1 Sol75.2 %
3. GPT-6 Astra74.8 %
4. GPT-6 Sol68.8 %

Source: Google blog post (Argon); OpenAI figures as tabulated by Vellum, Sept. 29, 2026. Vendor-supplied, not independently run.

The gap between Argon and OpenAI's GPT-6.1 Sol is 2.7 points, and each figure comes from the company selling the model.

Why the cyber group goes first

Google pitches Argon for finding, validating and patching software flaws. The post cites Wiz's Scan for Good initiative, which used the model to uncover a critical vulnerability in healthcare software. It also describes Argon converting up to 800,000 lines of C and C++ to Rust, including the Fuchsia Zircon kernel.

Google said it is taking a phased approach that includes the U.S. government's voluntary pre-release process. Its safeguards section lists misuse defence, prompt-injection resistance, chain-of-thought monitoring that can stop execution, and sandboxed environments.

That last item has a recent precedent. LeCun blamed leaky sandboxes for an agent breach at OpenAI and Hugging Face, a claim the investigators disputed. Google's post does not say how Argon's sandboxes were tested.

The model arrived the day after OpenAI published its GPT-6.1 Sol system card. Google said Argon outscored OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models, a claim TechCrunch reported as Google's own. Cloudflare's Clef release shows how differently an open-weight model publishes its figures.

Google has not said when Argon reaches the Gemini API for everyone. The Fairwind group is the only way in for now.