Google just went FRONTIER (Gemini 4 Argon)

Summary

Matthew Berman reviews Google’s Gemini 4 Argon, which he says he tested for about a week. His overall assessment is that Google is back among the leading AI model developers. Argon performs strongly across coding and knowledge-work tasks, though it is not his top choice for 3D graphics.

Coding and interactive demos

Berman tests Argon by asking it to build browser-based projects from relatively simple prompts:

• Cyberpunk 3D city in Three.js: The result features animated buildings, a flying car, visual effects, adjustable color themes, and camera settings. It reportedly runs at around 60 frames per second and took about 30 minutes to create. Berman notes that Argon often spends roughly that long on a task unless given a specific time limit or a /goal.
• Rube Goldberg machine in Three.js: The controllable simulation includes physics-driven sequences, camera modes, speed controls, and themed levels. Berman is impressed by the physics and overall result, but notices clipping. Repeated requests for hyperrealistic environments also fail to produce convincing realism.
• Fall Guys-style game: After several rounds of prompting, Argon produces a playable game with multiple levels, collisions, and an end-of-round animation. Berman finds it fun and says iteration substantially improves the final version.
• Ocean and boat scene in Unreal Engine: The boat and waves look good, though Berman notices some repetition in the water. He also notes that the scene does not let him go underwater.
• Excel clone: After about 20 minutes, Argon creates a browser-based spreadsheet with a convincing interface and working basic formulas and selection. A SUM operation initially causes an error, showing that the clone is not flawless.

Across the demos, Berman favors starting with broad prompts and then iterating rather than specifying every detail upfront. He also argues that software companies will need advantages beyond code that AI models can reproduce.

3D performance and model comparisons

Berman says Argon is better at 3D work than Meta’s newly released Muse Wonder, but still behind Astra and Opus. He describes Argon’s 3D capability as adequate rather than exceptional: it can generate impressive interactive scenes, but has difficulty with realism and occasionally has visual or collision issues. He suggests that browser-based Three.js may be a limiting factor compared with tools such as Unreal Engine or Blender.

Benchmarks and knowledge work

Berman reports that benchmarks shared by Box, which focuses on enterprise knowledge work, show a sizable gain over Gemini 3.1 Pro: Argon scores 85%, compared with 71%, on Box’s total dataset. The tests cover tasks such as spreadsheet evaluation, PDF extraction, presentations, images, and industry-specific document work. He also describes Box as a way for companies and developers to search and work across company documents using AI models.

He discusses additional benchmark results comparing Argon with other leading models:

• Deep Sweep: Argon scores 77.9%, behind Muse Wonder at 79%.
• Vibe Code Bench: Argon scores 91.9%, the highest among the models shown.
• Terminal Bench: Argon scores 57, below Opus at 66.4% and also behind Muse Wonder and Astra.
• OSWorld: Argon scores 69%, compared with GPT 6.1 Soul at 71.4%.
• Agent’s Last Exam: Argon scores 39, slightly above Opus at 38.

Berman concludes that Argon is competitive with leading models and says broader competition benefits consumers through better, cheaper, and faster models. He also remarks that Google’s release was unusual: benchmarks and model details were announced before the public could use it, while some testers and security researchers had early access.

Reviews, guides, and demonstrations

The video is primarily a hands-on model review, featuring coding demonstrations and benchmark comparisons. It does not provide a step-by-step tutorial, but it demonstrates prompt iteration and shows how the generated projects behave. Berman also briefly promotes Box as a tool for enterprise document search and AI-assisted knowledge work.

Main speaker and sources

• Main speaker: Matthew Berman, presenting the hands-on tests, opinions, and benchmark comparisons.
• Other sources mentioned: Box benchmark results and comments attributed to Meta AI head Alexander Wang about Muse Wonder’s lack of 3D-focused training.