Last updated: October 2026

What is Gemini 4 Argon, and can you use it yet?

Gemini 4 Argon is Google DeepMind's new frontier model, announced on September 30, 2026. Today it is limited to trusted cyber defenders through the Fairwind Program, with paid API customers and Google AI Ultra subscribers next and no date given, according to Google (2026).

This guide covers access, pricing, vendor-reported and independent benchmarks, and the security questions a CISO should ask before approving the model. Every figure is dated and linked to the page where it appears.

TL;DR: Key Takeaways

- Access is narrow. Gemini 4 Argon reached only Fairwind cyber defenders at launch, and DataCamp (2026) found no published API model ID on September 30.

- Price undercuts the rivals at launch. Google lists $2 per million input tokens and $10 per million output tokens, rising to $4 and $20, with a 95% discount on cached input (Google, 2026).

- Output is the headline spec. The output limit is 1 million tokens, up from 64,000, and Google has not stated an input context window (MarkTechPost, 2026).

- Benchmarks are mixed. Google reports 77.9% on DeepSWE v1.1 against 74.2% for Claude Opus 5.5, yet Opus 5.5 leads Terminal-Bench 4.0 at 66.4% versus 57.4%, and no third party had reproduced Google's table at launch (DataCamp, 2026).

- Independent boards disagree. Artificial Analysis scores Argon 52.6 against 57.6 for Opus 5.5, while Vals ranks it first of 41 at 68.9% (Trending Topics, 2026; Vals AI, 2026).

- Security claims need testing. Google says Argon leads Gray Swan's indirect prompt injection benchmark but published no score, so enterprises should run their own tests.

At a glance: Argon against Opus 5.5 and GPT-6 Astra

Sources: Google, DataCamp, VentureBeat, Trending Topics, MarkTechPost (output limits), NeuralTrust. Retrieved October 1, 2026.

What is Argon and who can use it today?

Gemini 4 Argon is a Google DeepMind frontier model announced on September 30, 2026 and released first to trusted cyber defenders through the Fairwind Program. Google says paid API customers and Google AI Ultra subscribers follow "as soon as possible", while the model also takes part in the U.S. government's voluntary pre-release review (Google, 2026).

Where Argon sits in the Gemini line

Google's last flagship series was Gemini 3 in November 2025, according to VentureBeat (2026). Fello AI (2026) adds that Argon is the first Gemini with a codename instead of the Pro, Flash and Flash-Lite tiers, and that Google has not said what other Gemini 4 models will be called.

Who has access right now

DataCamp (2026) reports that on launch day Argon was absent from OpenRouter, Vertex AI, Gemini CLI, Cursor and GitHub Copilot. Engineering teams cannot move production traffic to it yet. Security teams are the exception: Google says Wiz used Argon through its Scan for Good initiative and found a critical vulnerability in healthcare software that earlier frontier models had missed.

Argon pricing: what does it cost against Opus 5.5 and GPT-6 Astra?

During the introductory period Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens. Standard rates are $4 and $20, with a 95% discount on cached input. Google has not said how long the introductory price lasts. At standard rates Argon matches Claude Opus 5.5.

Sources: Fello AI (cached prices), VentureBeat (Opus 5.5 and Astra prices), NeuralTrust. The last column is our arithmetic from the listed rates.

Sticker price is only part of the bill. According to Trending Topics (2026), Artificial Analysis measured a cost per average task of $1.99 for Argon, $3.26 for GPT-6 Astra and $5.98 for Claude Opus 5.5. The same evaluation shows Argon used 110 million tokens to run the index against a median of 82 million, so verbosity eats into the saving. Artificial Analysis also notes that Argon's price is introductory. DataCamp (2026) adds that Google has not said how reasoning tokens are billed.

Gemini 4 Argon benchmarks: what Google reports

Google reports Gemini 4 Argon ahead on DeepSWE v1.1 at 77.9%, the Vals Index at 68.9%, AutomationBench at 51.3% and Harvey's Legal Agent benchmark, and behind on FrontierSWE v2 and Terminal-Bench 4.0. These are vendor-reported figures, and DataCamp (2026) notes that no third party had reproduced them at launch.

Source: DataCamp (2026), compiling Google's announcement. Retrieved October 1, 2026.

The pattern is specific. Argon leads on long-horizon software tasks, legal and finance work, long-context retrieval and business automation. MarkTechPost (2026) reports 51.3% on AutomationBench against 42.5% for Opus 5.5. It trails where agents work in a terminal or across a large codebase: a 10.5-point gap to GPT-6 Astra on FrontierSWE v2 and a 9-point gap to Opus 5.5 on Terminal-Bench 4.0. Google also reports 91.7% on LVBench for long video understanding.

How we compared

We did not run these benchmarks. We use vendor-reported numbers where they are the only source and label them as such. Where an independent board exists we show it beside the vendor figure. Prices come from the vendors or from a named outlet, and every table carries its retrieval date.

What independent evaluations say about Argon

Independent evaluations place Gemini 4 Argon at or near the top on some boards and behind Claude Opus 5.5 on others. Artificial Analysis scores it 52.6 against 57.6 for Opus 5.5, while Vals ranks it first of 41 models on its index at 68.9%. The honest summary is a strong, uneven model.

Sources: Trending Topics (2026) for Artificial Analysis, Vals AI (2026). Artificial Analysis tested only Argon's High setting. Retrieved October 1, 2026.

The Decoder (2026) reports a hallucination rate of 15% for Argon against 51% for GPT-6 Astra, but factual accuracy of 50% against 63%, and a first place on the Arena.ai Text Arena at 1,525 points. It concludes that Google closes the gap without taking a clear lead. Axios (2026) adds that Bloomberg reported some Google staff found performance lacking in internal testing, a claim Google disputed.

Context window and 1M-token output limit

Google states a 1 million token output limit for Gemini 4 Argon, up from 64,000 on earlier Gemini models, and does not state an input context window. Some outlets report a 1 million token input window, but we could not match that figure to a Google source, so treat it as unconfirmed.

The output figure is the real differentiator. MarkTechPost (2026) reports that Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra each cap a single response at 128,000 tokens. Google cites internal migrations of C and C++ code to Rust at up to 800,000 lines, including the Fuchsia Zircon kernel, as the target use case.

A response of that size has a cost and a risk. At $10 per million output tokens a maximal answer costs up to $10 at introductory pricing and $20 at standard pricing. More important, an agent that writes 800,000 lines in one pass produces a diff no human can review line by line. The output limit raises the stakes of every guardrail around the agent.

Gemini 4 Argon security: what enterprises need to know

Google reports that Gemini 4 Argon leads Gray Swan's indirect prompt injection benchmark and ships with chain-of-thought and action monitors that can stop execution. It published no injection score, and DataCamp (2026) reports that the Fairwind cohort received the model without cyber safeguards. Model-level defenses help, but they do not replace controls around the agent.

What Google says it built in

Google lists refusal training for harmful requests, safeguards for cyber and CBRN misuse under its Frontier Safety Framework, internal and external red teaming, and activation monitoring to detect misuse. It adds automated red teaming and adversarial training against prompt injection, plus monitors that watch reasoning and actions and can halt a run. DataCamp (2026) notes Google kept monitoring findings out of training so the model does not learn to evade them.

……（原文過長，此處截斷）