Gemini 4 Argon: No-Hype Full Review & Testing (Better than Opus?)

Summary
Pat Simmons conducts a blind comparative benchmark of Google DeepMind’s newly released Gemini 4 Argon against Anthropic’s Claude Opus 5.5, Claude Fable 5.1, and OpenAI’s GPT-6 Astra. Across four demanding generative coding and development tasks—an interactive 3D product render, a promotional launch video, an in-browser 3D game clone, and a complex website replication—he evaluates each model on visual fidelity, token consumption, build runtime, and financial cost.

---

What is shown

• Official Benchmarks & Demos Review [00:46–07:18]: Simmons examines DeepMind blog announcements and promotional demos, highlighting 3D Blender modeling (a multi-building pagoda complex, a tabletop racecar track, a raytraced room with Lyria 3.5 music and Gemini 3.8 TTS voiceover) and benchmark tables (DeepSWE v1.1, Harvey’s Legal Agent Benchmark, AutomationBench, Terminal-Bench 4.0, Gray Swan IPI).
• Test 1: Photoreal 3D Cold Brew Bottle [07:20–14:50]: Prompting models headless in Blender via Python scripts to create an interactive 3D product view with liquid-pouring mechanics for a fictional brand (“Clanker Coffee Co.”). Model D (Claude Opus 5.5) wins with realistic glass refractions and procedural pouring; Model B (Gemini 4 Argon) produces a flat, illustrative render with unnatural liquid animation and finishes last.
• Test 2: Launch Video for Clanker Coffee Co. [14:51–26:48]: Prompting agents to generate a full product commercial using Motion-Feel and ElevenLabs audio. Model C (Claude Opus 5.5) places first with cinematic lighting, smooth easing, and realistic motion blur; Model B (Gemini 4 Argon) ranks fourth due to cluttered, illegible on-screen text overlays and jarring pacing.
• Test 3: Browser Game Development (Rocket League Clone) [26:49–33:01]: Tasking models to code a playable 3D car soccer game with stadium lighting, driving mechanics, and physics. Model A (Claude Opus 5.5) achieves smooth wall-driving, responsive aerial physics, and goal effects; Model B (Gemini 4 Argon) finishes last due to collision clipping, glitchy goal posts, and poor car handling.
• Test 4: Pixel-for-Pixel Website Clone (Oryzo.ai) [33:02–43:14]: Replicating an interactive 3D scrollytelling site featuring a cork coaster, 3D hand animation, and dynamic shaders. Model B (Claude Opus 5.5) wins with accurate procedural textures and seamless transitions; Model C (GPT-6 Astra) successfully models an articulated 3D hand but took 499 minutes; Model A (Gemini 4 Argon) places last due to missing coffee cup 3D assets and distorted textures.
• Final Verdict [43:15–45:21]: Simmons summarizes that while Gemini 4 Argon was the cheapest and fastest model in every test, its creative and 3D output quality consistently fell behind its rivals, concluding it did not match Google’s launch demos.

---

Claims & numbers

• Gemini 4 Argon Pricing: Promotional launch price through November 30, 2026, is $2.00 per million input tokens and $10.00 per million output tokens (with cached tokens at 75% input discount), rising to $4.00/$20.00 (matching Claude Opus 5.5 list price) after the promotion [03:15].
• Harvey’s Legal Agent Benchmark: Google claims Gemini 4 Argon scores 19.6%, compared to 6.7% for Claude Fable 5.1, 5.4% for GPT-6 Astra, and 3.8% for Claude Opus 5.5 [00:06, 04:14].
• AutomationBench: Gemini 4 Argon reports 51.3%, vs. 42.5% for Claude Opus 5.5, 41.4% for GPT-6 Astra, and 31.4% for Claude Fable 5.1 [00:10, 04:22].
• DeepSWE v1.1: Gemini 4 Argon claims 77.9%, compared to 74.2% for Claude Opus 5.5, 74.1% for GPT-6 Astra, and 67.4% for Claude Fable 5.1 [00:11, 04:54].
• Terminal-Bench 4.0: Gemini 4 Argon scores 57.4%, trailing Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2% [05:07].
• Gray Swan IPI Benchmark: Google claims Gemini 4 Argon achieves a 0.7% attack success rate [06:43].
• Test 1 Run Statistics:
  • Gemini 4 Argon: $6.85, 9.68M input / 383.8K output tokens, 36 min build time (Rank 4) [13:48].
  • GPT-6 Astra: $8.99, 5.06M input / 46.5K output tokens, 34 min build time (Rank 2) [13:56].
  • Claude Fable 5.1: $17.96, 18.19M input / 155.6K output tokens, 57 min build time (Rank 3) [14:02].
  • Claude Opus 5.5: $29.95, 77.29M input / 330.4K output tokens, 114 min build time (Rank 1) [14:09].
• Test 2 Run Statistics:
  • Gemini 4 Argon: $10.14, 17.77M input / 438.5K output tokens, 80 min build time (Rank 4) [25:16].
  • GPT-6 Astra: $28.44, 19.72M input / 83.7K output tokens, 81 min build time (Rank 2) [25:22].
  • Claude Fable 5.1: $54.26, 62.12M input / 242.5K output tokens, 302 min build time (Rank 3) [25:29].
  • Claude Opus 5.5: $44.98, 109.38M input / 313.3K output tokens, 318 min build time (Rank 1) [26:00].
• Test 3 Run Statistics:
  • Gemini 4 Argon: $10.59, 20.94M input / 479.2K output tokens, 54 min build time (Rank 4) [32:04].
  • GPT-6 Astra: $14.70, 8.96M input / 79.5K output tokens, 56 min build time (Rank 3) [32:07].
  • Claude Fable 5.1: $32.30, 19.99M input / 207.7K output tokens, 77 min build time (Rank 2) [32:10].
  • Claude Opus 5.5: $46.24, 143.08M input / 407.6K output tokens, 155 min build time (Rank 1) [32:14].
• Test 4 Run Statistics:
  • Gemini 4 Argon: $11.95, 29.56M input / 320.5K output tokens, 77 min build time (Rank 4) [42:21].
  • Claude Opus 5.5: $67.87, 189.14M input / 467.2K output tokens, 129 min build time (Rank 1) [42:24].
  • GPT-6 Astra: $357.84, 271.95M input / 953.3K output tokens, 499 min build time (Rank 2) [42:30].
  • Claude Fable 5.1: $98.54, 78.72M input / 679.2K output tokens, 112 min build time (Rank 3) [43:12].
• Subscription Usage: The presenter notes he hit daily usage caps on Google’s $99/month subscription tier after approximately one full build [26:19].

---

Notable quotes

• “So is Gemini 4 the real deal, or is it just another benchmaxer? Let’s find out.” [00:14]
• “Hey Google, what happened to your pagoda here? ...Because this 3D work and this 3D work are much different.” [12:57]
• “Opus 5.5, I’m telling you, Opus 5.5 is on a generational run right now with these model tests.” [24:23]

---

Assessment
This is an authentic, independent review and blind side-by-side benchmark test by tech creator Pat Simmons. The testing methodology is rigorous and fully executed within live headless browser/Blender evaluation harnesses, demonstrating that despite strong official synthetic benchmark scores, Gemini 4 Argon struggled with complex spatial 3D generation and visual design tasks compared to current rival frontier models.

Described by gemini-3.8-flash on 2026-10-15 from the video’s audio and frames.