I tried Gemini 4 Argon | Physical product building with Gemini 4

Summary

In this video, Dhaval Patel (founder of codebasics) is joined by guest Siddhant Pandey (Google Developer Expert and AI Research Engineer) to review practical real-world engineering and coding capabilities of Google DeepMind’s newly released frontier model, Gemini 4 Argon. They walk through six complex technical projects generated with Gemini 4 Argon, focusing heavily on physical product CAD/specification modeling, generative 3D simulation interfaces, textbooks, and interactive technical video generation.

---

What is shown

• 00:00 - 01:14: High-level teaser of six generated projects: an engineered summer motorcycle helmet digital twin, a phone jail “Focus Desk Organizer”, a 500-page Generative AI Python textbook, a 1:1 digital twin of the International Space Station (ISS), a gamified Solo Leveling-themed fitness web application, and a 3Blue1Brown-style GPU architecture explainer video.
• 01:16 - 02:05: Discussion of Google’s official launch blog post for Gemini 4 Argon and access options (Gemini web app, Google AI Studio, and Google Antigravity).
• 02:06 - 05:48: Project 1: Budget Summer Helmet Digital Twin:
  • Interactive 3D web UI (Three.js/WebGL) showing exploded views, airflow ventilation lines, and colorway toggles [02:08, 03:20, 03:38, 03:54].
  • Detailed engineering requirements and technical specifications PDF generated by Gemini 4 Argon according to Indian BIS certification (IS 4151) [04:35 - 05:45].
• 05:48 - 09:35: Review of the Python automation script used with Blender in headless mode to generate .blend files, dimensioned orthographic drawings, and CAD STL files for each individual component [05:48, 07:50 - 08:04, 09:15 - 09:27].
• 10:04 - 10:23: Review of Rapid prototyping suppliers (such as Think3D and PrintForge in India) suggested by Gemini 4 Argon for physically 3D-printing the parts.
• 12:25 - 13:28: Project 2: Gamified Solo Leveling Fitness App:
  • Siddhant Pandey demonstrates a customized wellness web application inspired by the anime Solo Leveling, featuring daily quests, habit tracking, and workout logs built using Gemini 4 Argon [12:25 - 13:28].
• 15:12 - 19:20: Project 3: “Python for Gen AI: The Great Indian Kitchen Edition”:
  • A 500-page textbook PDF explaining core Python concepts (data types, string manipulation, LLM pipelines) through Indian food analogies (e.g., using a Masala Dabba model for memory, clove counts for integers, liquid ghee for floats) [15:12, 16:40 - 17:08, 18:16 - 19:11].
• 20:06 - 25:32: Project 4: “Focus Desk Organizer” (Phone Jail Box):
  • Interactive 3D model of a lockable distraction-free desk organizer with an integrated reading lamp, sliding calculator drawer, rotary timer, and simulated LED countdown progress display [20:08, 20:47, 22:06 - 22:23, 24:55 - 25:30].
• 25:50 - 31:16: Project 5: ISS 1:1 Orbital Digital Twin:
  • Siddhant demonstrates a complete, interactive, modular 3D model of the International Space Station, complete with sun-tracking solar arrays, radiator steering, exploded assembly phases, and automated camera presets [25:51 - 27:30, 28:20 - 28:44].
• 31:24 - 36:06: Project 6: 3Blue1Brown-style GPU Architecture Video:
  • Siddhant plays a video created using Python and the Manim library with AI-synthesized narration and synchronized animations explaining GPU matrix multiplication, CPU vs. GPU architecture, and CUDA kernel thread indexing [31:24, 32:25, 33:45 - 34:02, 34:45 - 35:55].

---

Claims & numbers

• 00:23 - 00:29: Benchmark table shows Gemini 4 Argon scoring:
  • 68.9% on Vals Index (vs. 63.1% GPT-6 Astra, 65.8% Claude Fable 5.1).
  • 51.3% on AutomationBench (vs. 41.4% GPT-6 Astra, 31.4% Claude Fable 5.1).
  • 65.4% on Vals Finance Agent v2 (vs. 53.5% GPT-6 Astra, 58.9% Claude Fable 5.1).
  • 19.6% on Harvey’s Legal Agent Benchmark (vs. 5.4% GPT-6 Astra, 6.7% Claude Fable 5.1).
  • 77.9% on DeepSWE v1.1 (vs. 74.1% GPT-6 Astra, 67.4% Claude Fable 5.1).
  • 55.0% on FrontierSWE v2 (trails GPT-6 Astra’s 65.5%).
  • 91.9% on Vibe Code Bench.
• 00:43 - 00:47: Dhaval states Gemini 4 Argon has a single-pass output limit cap of up to 1,000,000 tokens (compared to 64K tokens for previous Gemini models).
• 07:15 - 07:18: Dhaval notes the initial Blender script generation run took approximately 1 hour before requiring a restart and refinement.
• 11:18 - 11:21: Siddhant mentions a decent desktop 3D printer in India costs approximately ₹30,000 to ₹50,000.
• 16:06 - 16:13: Generating the complete Python book draft took 17 minutes for the first pass and approximately 2 to 3 hours total to expand to 500 pages.
• 28:10 - 28:19: Siddhant states generating the full multi-phase ISS Blender model and code took about 2 to 3 hours.
• 33:04 - 33:13: Generating the complete Manim-based GPU explainer video took roughly 1 hour of model compute time and about 15 minutes of human prompting time.

---

Notable quotes

• 00:44: “Unprecedented 1 million token output limit in a single pass.” — Dhaval Patel
• 21:59: “This is Three.js visualization. I did it the same way we did the helmet demo.” — Dhaval Patel
• 33:33: “It is using Manim Community... this was without any kind of pictorial references or any kind of links that I provided. This was all Argon.” — Siddhant Pandey

---

Assessment

This is a technical hands-on review and demonstration by two developers testing Google’s Gemini 4 Argon frontier model. The interactive Three.js web dashboards, Blender models, code editors, and generated outputs are genuine working code artifacts running in local development environments (localhost). The presenters are transparent about generation times, retries, and errors encountered during multi-hour code generation runs.

Described by gemini-3.8-flash on 2026-10-13 from the video’s audio and frames.