Google's Gemini 4 Argon Achieves Frontier Status, Demonstrates Strong 3D and Software Development Capabilities

Google has launched Gemini 4 Argon, a new AI model that Matthew Berman, an early tester, states has brought Google up to the current AI frontier. Berman, who had access to the model for about a week, highlighted its impressive capabilities.

The model demonstrated strong performance in browser-based 3D rendering, which was a focus of Berman's initial testing and evaluations.

During testing, Gemini 4 Argon was tasked with creating a hyperrealistic cyberpunk 3D city scene using 3.js. The model successfully generated the complex animation, which ran smoothly at 60 frames per second directly in the browser.

The generated city scene featured extensive details, including numerous 3D buildings, flying cars, and a significant amount of neon colors. Users could customize elements such as neon bloom density, sky traffic, and camera controls, including '2049 amber', 'Biosynth', and 'Red Syndicate' modes.

The entire generation process took approximately 30 minutes, using a very simple initial prompt. Berman noted that Argon, when not given specific time constraints, tends to spend around 30 minutes on a given task.

Meta's recently released Muse Wonder, also a frontier model, struggles with 3D tasks, including those involving 3.js, Unreal Engine, and Blender. Alexander Wang, head of AI at Meta, acknowledged that the team either forgot or failed to apply Reinforcement Learning (RL) training for 3D tasks, considering it a retrospective mistake.

In comparison, Argon performs better than Muse Wonder in 3D rendering. However, Berman noted that Argon is not as proficient in 3D as other models like Astra and Opus, which he described as "groundbreaking" for 3D applications. Despite this, Argon shows high proficiency in general coding tasks beyond 3D.

Gemini 4 Argon successfully constructed a 3D Rube Goldberg simulation that was fully controllable and featured realistic physics. The simulation included interactive components where users could manually trigger elements, such as a ball to start the chain reaction.

The Rube Goldberg machine included various themed levels, such as a Grand Canyon version, a Las Vegas version, a clockwork lab, and a Kyoto temple. While some minor clipping issues were observed, the overall physical interactions were rated as very good. Users could also adjust camera modes (auto-orbit, chase, wide shot) and control the simulation speed.

Despite being prompted to create hyperrealistic levels, Gemini 4 Argon struggled to achieve this aesthetic within the 3.js framework. The thematic levels, such as the Grand Canyon or Las Vegas, did not appear highly realistic.

The creator suggested that this limitation might stem from 3.js itself, especially when operating within a browser environment. It is possible that developing similar projects in Unreal Engine or Blender could yield significantly better visual fidelity.

According to benchmarks provided by Box, Gemini 4 Argon shows a substantial performance increase compared to its predecessor, Gemini 3.1 Pro. The total dataset benchmarks indicate a 14% jump in performance, improving from 71% to 85%.

These improvements span across various industry-specific tasks crucial for knowledge work, including spreadsheet evaluation, PDF extraction, presentation analysis, and image processing. Box AI integrates models like Argon to enhance enterprise document search capabilities, with the speaker's company, Forward Future, being a user of the service.

Gemini 4 Argon successfully developed a playable game level resembling a 'Fall Guys' clone, complete with collision mechanics and engaging gameplay. Initially, the graphics experienced flickering, but these issues improved during the testing period.

The creator adopted a strategy of using simple prompts to avoid over-specification, pushing the model to generate up to 10 complex levels. Iterative prompting was key, with instructions like "Make the levels even bigger and more complex" and "Look out for collisions that aren't being calculated, balls rolling down, for example," leading to significant improvements in the game's complexity and physics.

When compared to other frontier models, Gemini 4 Argon demonstrated superior 3D rendering capabilities compared to Meta's Muse Wonder. However, it still falls short of the specialized 3D quality achieved by models like Opus and Astra.

Testing in Unreal Engine, specifically with a boat floating in the ocean, revealed good wave animation and high detail in the water. Despite this, the model had limitations, such as not allowing users to go underwater, a feature present in some other models.

To assess its real-world software engineering capabilities, Gemini 4 Argon was given the task of cloning Microsoft Excel. Without a pre-set loop, the model ran for approximately 20 minutes, during which it successfully opened Excel locally and created a functional, browser-based user interface.

The cloned Excel demonstrated basic arithmetic functions, such as 1+2, and successfully executed the 'sum' function by allowing users to highlight rows for calculation. These functionalities underscore the model's proficiency in developing complex software applications.

The creator suggested that this capability indicates a significant future reduction in software production costs. He argued that if a product's sole competitive advantage lies in its lines of code, it will face challenges, emphasizing the need for unique products and services that cannot be easily replicated by advanced AI models.

The rollout of Argon involved a unique closed testing period for early testers and security researchers before its public release. This allowed for comprehensive evaluation against other industry-leading models.

A custom comparison chart, compiled by the speaker, included models such as Muse Wonderful, GPT 6.1 Soul, Astra Fable 5.1, and Opus 5.5. On the Deep Sweep benchmark, which measures engineers' perception and usage, Argon scored 77.9%, placing it just behind Muse Wonderful, which led with 79%.

Argon achieved the top position on the Vibe Code Bench, scoring 91.9%. However, in the Terminal Bench, which assesses a model's ability to write and execute commands, Argon scored 57, trailing Opus 5.5 (66.4%), Muse, and Astra. Despite not being the leader in all categories, its performance was considered strong and competitive.

In additional benchmarks, Argon scored 69% on OS World, a test of a model's ability to control a computer, performing similarly to GPT 6.1 Soul at 71.4%. On the Agent/Exam benchmark, Argon achieved a score of 39%, slightly surpassing Opus, which scored 38%.

The speaker highlighted that Google's entry with Argon means that major tech companies, including Google, Meta, OpenAI, and Anthropic, now possess competing frontier AI models, with Apple and Microsoft being notable exceptions. This heightened competition is expected to benefit consumers through lower prices, faster speeds, and improved model quality.