Gemini 4 Argon: Google’s Frontier Model for Coding, Knowledge Work, and Cyber Defense

Gemini 4 Argon is Google DeepMind’s first major Gemini 4 model. Announced on 30 September 2026, it became widely available on 13 October 2026. It is built for complex, multi-step work in software engineering, legal, finance, cybersecurity, and science.

I have tracked every major model release this year. Argon marked Google’s clear return to the frontier after a quieter period. It delivers strong results on real-world tasks and comes at a competitive price.

What Gemini 4 Argon Is

Argon is a frontier multimodal model. It accepts text, images, video, and audio. It supports a 1 million token context window shared between input and output.

Google designed it for long-horizon work. That means the model can stay focused on difficult problems that take many steps. It reduces hedging, repetition, and generic language compared with earlier Gemini models.

Thousands of Google employees already use it for specialized coding, research, and writing tasks.

Key Benchmark Results

Google published a detailed comparison table. Argon led or tied for the lead on most rows. Independent tests place it in the top tier but not always first.

Benchmark	What It Measures	Argon Score
Vals Index	Knowledge work (finance, legal, coding, tax) weighted by GDP	68.9% (#1)
DeepSWE v1.1	Long-horizon real-world software engineering	77.9%
AutomationBench	Business automation tasks	51.3%
Vals Finance Agent v2	Finance agent performance	65.4%
Harvey’s Legal Agent Benchmark	Legal agent work	19.6%
CWE-bench v1	Vulnerability remediation	68.0% (tie)

On the independent Artificial Analysis Intelligence Index (earlier version), Argon scored 53 at high reasoning. That tied GPT-6 Astra and sat a few points behind Claude Opus 5.5.

Argon also showed a notably low hallucination rate on AA-Omniscience tests compared with several rival models.

Strengths Across Domains

Argon performs especially well in three areas.

Knowledge work — It leads the Vals Index and shows clear gains in legal and finance agent benchmarks.
Long-horizon coding — DeepSWE score of 77.9% highlights ability to handle multi-step engineering problems.
Cybersecurity — Strong results on vulnerability remediation tasks.

It also improved long-video understanding and long-context graph reasoning compared with previous Gemini releases.

Pricing and Availability

Introductory pricing runs through 30 November 2026:

Token Type	Introductory Price	After 30 Nov 2026
Input (per 1M)	$2.00	$4.00
Output (per 1M)	$10.00	$20.00
Cached input	95% off	95% off

Access is now broad for paid users. It is available on the paid Gemini API, Google AI Studio, Gemini for business, the Gemini app (Pro and Ultra subscribers), AI Mode in Search, and Antigravity for Ultra subscribers. Free-tier users cannot access it.

The model ID is gemini-4-argon.

How Argon Compares to Later Models

Two weeks after Argon’s wide release, Google announced Gemini 4 Carbon. Carbon improved further on cybersecurity and long-horizon tasks and took the top spot on the updated Artificial Analysis Intelligence Index v5.0 with a score of 59.

Argon remains the widely available Gemini 4 model for most users. Carbon started with limited Fairwind access, the same phased approach Argon used at launch.

My Assessment

Argon is a solid, practical frontier model. It is especially useful for teams that need strong performance on knowledge work, legal and finance tasks, and multi-step coding. The low introductory price and 1 million token context make it attractive for heavy workloads.

Independent rankings show it is competitive but not always the absolute leader. Google’s own tables are more favorable. Real-world use by early teams and Google’s internal adoption give more confidence than pure leaderboard scores.

The phased rollout — first cyber defenders, then paid users — followed a careful pattern that Google continued with Carbon.

Frequently Asked Questions

Is Gemini 4 Argon still the best Google model available to most users?

Yes. Carbon has higher benchmark scores in several areas but remains limited to selected cyber defenders. Argon is the model most developers and subscribers can use today.

How long does the $2 / $10 pricing last?

Until 30 November 2026. After that the rates rise to $4 input and $20 output per million tokens.

What is the context window?

1 million tokens shared between input and output.

Does Argon support multimodal inputs?

Yes. It accepts text, images, video, and audio.

Looking Ahead

Argon established Google’s presence at the frontier again. Carbon shows the company is moving quickly with further improvements, especially in security and long-running agentic work.

For teams that need a capable, widely available model right now at a competitive price, Argon remains a strong choice. Watch how Carbon expands access and how independent evaluations evolve over the coming months.