Google Gemini 4 Argon: Benchmarks, Pricing, and Real-World Performance

Google has released Gemini 4 Argon, a new frontier model trained from scratch over the last six months using Google's TPUs and the JAX Python library. This release comes after the previous Gemini 3.5 Pro failed to meet internal benchmarks and was not released. Argon is positioned as a powerful thinking sand, aiming to surpass competitors like GPT-6 Astra and Claude Opus 5.5 in various benchmarks while being more cost-effective and faster.

Naming and Availability

The name "Argon" is an interesting choice, translating to "lazy" in Greek. However, it's suggested to have a hidden meaning, implying "your tokens are gone." Google initially announced Argon on September 30th, making it available only to a select group of "trusted cyber warriors" through the Fairwind program, who had access without guardrails. For the general public, the model is now available with guardrails, and it's expected to produce significantly fewer refusals than previous models.

Technical Architecture and Training

Argon is not a fine-tuned version of older Gemini models. It is a sparse mixture of experts transformer, trained from scratch on Google's TPUs. Its knowledge cutoff is March 2026, indicating at least six months of training.

Performance Benchmarks

Argon shows impressive performance in several benchmarks:

• Deep Software Engineering Benchmark: State-of-the-art for long-horizon software engineering tasks.
• Val Index: Ranks number one, weighing topics like finance, legal, and coding based on their share of the US GDP.
• AI Lawyer: Gemini is now considered the best AI lawyer.
• Long Video (LVbench): Dominates with 91.7%.

However, Argon does not win in all categories:

• Terminal Bench 4.0: Opus 5.5 wins 66 to 57.
• Frontier S: GPT-6 Astra wins by more than 10 points.
• Arena AI (Text Benchmark): Gemini Argon comes out on top.
• Code Arena (Webdev): Drops to number nine, suggesting it may not be as strong as Claude or GPT-6 for coding tasks.

Pricing and Developer Launch Issues

Until November 30th, Argon is priced at $2 per million input tokens and $10 per million output tokens, which is considered a competitive price for its capabilities.

Despite the competitive pricing, Google faced issues with the developer launch. Gemini 3.1 was replaced with Argon, but initially, Argon was only available to users with the AI Ultra plan, which costs $100 per month. This angered many "pro users" who only gained access a few days later with a smaller quota.

Real-World Demos and Comparisons

Video Generation

One of Gemini's selling points is its "omni model" capability, allowing it to take various input forms (image, video, sound) and output them. While technically ChatGPT and Claude cannot output raw video, Claude has become proficient at generating motion graphics by writing code that produces the sequence, offering more customization and analysis.

When tasked with generating a promo video for an app, Argon produced a "mid" result that was clearly AI-generated and deemed unsuitable for a high-quality app. In contrast, Claude, using its code-generation approach, produced a much better-looking and more customizable video.

3D Graphics Rendering (3JS)

A challenging test for frontier models is rendering realistic interactive 3D graphics in 3JS. Argon, Astra, and Opus were all prompted to "create a realistic interactive 3D horse with 3JS."

• Claude and GPT-6: Produced impressive results with some minor flaws that could likely be fixed with additional prompts.
• Gemini: The result was significantly different. While the UI was clean and expressive, the 3D model of the horse was low-poly, despite the prompt asking for a "realistic" horse. More significantly, Gemini's model included a "copulatory organ" that was disproportionately large and out of alignment, raising concerns about safety and appropriate content generation.

Macroscope: AI for Code Review

The video is sponsored by Macroscope, an AI-powered tool designed to help developers manage the influx of pull requests from coding agents. Macroscope builds its own AST walkers for common programming languages to provide real context for code changes, making reviews faster, cheaper, and more precise. It also runs custom team rules as native GitHub checks, blocking failing pull requests. If a PR passes these checks, a code review, and has a small blast radius, Macroscope can auto-approve it without human intervention. The service is currently auto-approving 40% of its customers' pull requests.