Summary
Matthew Berman puts Google’s Gemini 4 Argon through its paces after about a week of access, arguing that Google has caught up to the frontier. He tests the model with deliberately simple prompts and shows what it builds, including a cyberpunk city in Three.js running at 60 frames per second, a controllable 3D Rube Goldberg machine, and a Fall Guys-style game refined over several iterations.
Berman notes that Argon tends to work on tasks for around 30 minutes unless told otherwise or given a goal. He finds its 3D work good but not on par with Opus and Astra, while clearly ahead of Meta’s newly released Muse Wonder, which Alexander Wang has acknowledged was not trained to be strong at 3D. He also points out limits, including clipping issues and an inability to reach truly hyperrealistic results in the browser.
On benchmarks, partner Box reports that Argon scored 85% on its knowledge-work dataset covering data extraction and document analysis across legal, life sciences, and healthcare, up from 71% for Gemini 3.1 Pro. Viewers get a practical sense of the model’s strengths in coding and creative 3D, where it still falls short, and how it compares to rival frontier models.
📺 Source: Matthew Berman · Published October 14, 2026
🏷️ Format: Review
