Reviews

Analysis: Gemini 4 Argon — Google's first Gemini 4 model, its introductory pricing, and who should use it

Google has opened Gemini 4 Argon to developers and paying subscribers after a cyber-defender-only first phase. We look at the specs, the $2/$10 introductory price that doubles on 1 December, Google's benchmark claims and who it suits.

Abstract illustration of flowing streamlines tracing a vector field, in deep blue and cyan.
Abstract illustration of flowing streamlines tracing a vector field, in deep blue and cyan.

Google made Gemini 4 Argon generally available on 13 October 2026, opening it to developers on the paid tier of the Gemini API, to enterprises, and to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. The release comes two weeks after Google first announced the model on 30 September and limited it to a group of cybersecurity defenders. Google describes Argon as a frontier model for real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defence.

This analysis is based on Google’s published announcements, model card, pricing page and developer documentation, and on benchmark results reported by Google; we have not tested the model hands-on.

A phased release

Argon is the first Gemini 4 model, and Google chose not to release it to everyone at once. “Safely releasing frontier capabilities at this level requires a phased approach,” the company wrote in its 30 September announcement. The first phase went to trusted cyber defenders through Google’s Fairwind Program, and for those partners Google said it would release Argon “without cyber guardrails.” Google also said it was taking part in the US government’s voluntary process for pre-release model access.

The 13 October release is the second phase. Alongside general availability, Google said it was expanding Fairwind to hundreds of additional partners, and that about 10,000 scientists could apply for six months of free access to the highest usage tier of Antigravity, its agent-based development platform.

Specs

Gemini 4 Argon
API model ID gemini-4-argon
Architecture Sparse mixture-of-experts transformer; Google says it is not a fine-tune of a prior model
Inputs Text, images, audio, video and PDF
Output Text
Context window 1,048,576 tokens, shared between input and output
Maximum output 65,536 tokens by default; up to the full window with “long decode”
Thinking levels Low, medium, high
Knowledge cutoff March 2026

The output limit is the most unusual line. Google says it has raised the model’s output token limit to 1M tokens, “up from the previous 64K tokens.” That is aimed at very long jobs. Google’s own example is migrating C and C++ code to Rust, including up to 800,000-plus lines for its Fuchsia Zircon kernel.

Pricing

Gemini API, Standard tier (per million tokens) Until 30 November 2026 From 1 December 2026
Input $2.00 $4.00
Output (including thinking tokens) $10.00 $20.00
Cached input (implicit caching only) $0.10 $0.20

Argon is not available on the Gemini API’s free tier. Google calls the current rate an introductory price, and it doubles on 1 December. For comparison, Anthropic lists Claude Opus 5.5 at $4 input and $20 output per million tokens, the same as Argon’s post-introductory rate.

What Google’s benchmarks show

Google’s model card compares Argon with OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Claude Opus 5.5. All figures below are Google-reported. The card does not state the settings used for each model or where the competitor scores came from.

Benchmark (Google-reported) Gemini 4 Argon GPT-6 Astra Claude Fable 5.1 Claude Opus 5.5
Vals Index 68.9% 63.1% 65.8% 67.0%
Vals Finance Agent v2 65.4% 53.5% 58.9% 58.6%
Harvey’s Legal Agent Benchmark 19.6% 5.4% 6.7% 3.8%
AutomationBench 51.3% 41.4% 31.4% 42.5%
DeepSWE v1.1 77.9% 74.1% 67.4% 74.2%
FrontierSWE v2 55.0% 65.5% 56.3% 62.3%
Terminal-bench 4.0 57.4% 58.2% 57.9% 66.4%
GraphWalks, 256K to 1M (BFS, F1) 84.2% 71.8% 65.0% 66.8%
LVBench (video) 91.7% 87.5% 79.7% 83.7%
CWE-bench v1 68.0% 68.0% 58.0% 67.0%

The pattern in Google’s own figures is clear. Argon leads in its table on enterprise knowledge work, long-context retrieval and video understanding. On coding the picture is mixed: it leads on DeepSWE but trails GPT-6 Astra and Claude Opus 5.5 on FrontierSWE v2 and Terminal-bench 4.0. Google also reports that Argon leads Gray Swan’s indirect prompt injection benchmark, with an attack success rate of 0.7% across 15 adaptive attempts.

Safety

Google assesses its models against Critical Capability Levels (CCLs) under its Frontier Safety Framework. For cybersecurity, the model card states: “We cannot rule out that Gemini 4 Argon has reached this CCL.” It says the same of the biological CCL. In both cases, Google says it developed a safety case and assesses the risk, with safeguards in place, as below the critical level. Critical levels for harmful manipulation and machine-learning R&D were not reached. Google also describes an automated monitor that checks agent reasoning and actions in real time, and says Argon was trained to pause for human confirmation before irreversible steps.

Who it’s for

A good fit: enterprises working on finance, legal and other knowledge-heavy agent tasks, where Google’s figures show its largest leads. Teams with very long inputs or outputs, such as large code migrations, will also want to look at the long-context results and the 1M-token output option.

Worth comparing first: developers whose agents spend most of their time in a terminal or on the coding tasks measured by FrontierSWE. Google’s own table shows rivals ahead there.

Budget planners: any cost model built on the introductory price needs revising for December, when input and output rates double.

What it means

Our take: Argon shows how frontier launches are changing. The model was released in stages, with cyber defenders first, participation in a government pre-release process, and a model card in which Google says it cannot rule out the model reaching its critical cyber and biological capability levels, relying instead on safeguards and a safety case. That context is now part of what customers are buying, alongside the benchmarks.

On the numbers Google has chosen to publish, Argon is a strong option for enterprise knowledge work and long-context tasks rather than a clear leader on every coding test. The introductory price makes this a sensible moment to run comparisons, provided teams plan for the December increase.

Sources

See All

More on Google

See All

Type to search. ↑ ↓ to navigate, ↵ to open.