I Tested Gemini 4 vs Opus 5.5 vs GPT-6 Astra vs Muse Wonder (5 Use Cases)

Summary

Peter Yang compares four frontier AI models—Anthropic’s Claude Opus 5.5, OpenAI’s GPT-6 Astra, Google’s Gemini 4 (Argon), and Meta’s Muse Wonder—across five diverse use cases. The tests evaluate each model’s coding and generative capabilities across 3D modeling, browser game creation, animated video storytelling, website design, and long-form writing/research synthesis.

What is shown

• Introduction & Overview [00:00 - 00:46]: Peter Yang introduces the comparison between Claude Opus 5.5, GPT-6 Astra, Gemini 4, and Muse Wonder following the recent launch of Gemini 4 and Muse Wonder.
• Test 1: 3D Blender Scene [00:47 - 03:51]:
  • Prompt: “Build Himeji Castle in Blender and render a 30-second cinematic at dawn in cherry blossom season. Get the architecture right and add music you compose in code.”
  • Opus 5.5 [01:12]: Renders detailed cherry blossom close-ups with falling petals and an accurate multi-tiered Himeji Castle model with traditional music.
  • Gemini 4 [01:51]: Renders dynamic camera cuts (including under a bridge), typographic title cards (“The White Heron Awakens”), and traditional flute instrumentation.
  • GPT-6 Astra [02:28]: Generates a reflective water moat scene with piano/orchestral music, though the castle geometry is slightly flatter.
  • Muse Wonder [03:03]: Renders a low-poly stylized version with basic textures and birds.
• Test 2: Browser Rail Shooter Game [03:52 - 07:02]:
  • Prompt: “Build a Star Fox-style rail shooter in three.js, with the ship, enemies, boss, and terrain modeled in Blender. Include a barrel roll, lasers, a boss fight, and a soundtrack.” Follow-up: “Switch up the level to Terrans against the Zerg, and model all the Zerg properly in Blender. The Terran ship can be a Wraith.”
  • Opus 5.5 (“Hellhound Squadron”) [04:30]: Playable Three.js game with barrel rolls, cockpit voice comms, boss fights, and bloom/lens flare effects.
  • Gemini 4 (“Terran vs. Zerg”) [05:11]: Playable game with extensive controls, first-person cockpit toggle (V), Yamato Cannon ability (B), HUD counters, and multiple camera views.
  • GPT-6 Astra (“Iron Veil”) [06:00]: Simpler arcade rail-shooter aesthetic with basic enemy types and lighting.
  • Muse Wonder [06:38]: Basic 3D track shooter with rudimentary shapes and flying craft.
• Test 3: Animated Emotional Video Short [07:03 - 11:19]:
  • Prompt: “Make a 30-second animated short that tells a story and leaves viewers with a strong emotion. Make every frame and sound in code, with no stock assets.”
  • Opus 5.5 [07:27]: 3D animated cabin through seasons with counter (“Night 3,651”) where a woman waits with a porch light on for a returning car and reunites with her partner.
  • Gemini 4 (“Paper Wings”) [08:19]: Stylized motion graphic of glowing origami birds crossing distances, displaying excuses (“Tomorrow”, “Next Month”), ending with “Call home tonight.”
  • GPT-6 Astra (“One Small Light”) [09:18]: 2D animation of a small robot watering a potted plant inside a dark greenhouse until it grows into a tree.
  • Muse Wonder [10:22]: 2D story of an old man sheltering a flower from rain and nursing it, ultimately placing it on a tombstone next to a lighthouse.
• Test 4: Website Brand & Homepage Redesign [11:20 - 14:15]:
  • Prompt: “Redesign the brand and homepage for Behind the Craft, my newsletter, YouTube channel, and podcast. Create a logo, a color and type system, and a homepage that doesn’t look AI-generated. Don’t look at my current site.”
  • Opus 5.5 [11:49]: Editorial newspaper/magazine layout with interactive folded-corner graphic and brand guidelines sheet.
  • Gemini 4 [12:20]: Feature-dense terminal/technical documentation layout accompanied by a full Swiss design system breakdown.
  • GPT-6 Astra [13:06]: Bold cobalt blue and yellow high-contrast design featuring integrated generated photography and custom vector illustrations.
  • Muse Wonder [13:43]: Text-heavy serif publication layout resembling a traditional blog/newspaper index.
• Test 5: Writing & Medical Synthesis Evaluation [14:16 - 17:46]:
  • Yang tests models on converting raw interview transcripts into structured newsletter takeaways, as well as complex oncology research synthesis (evaluating clinical trial options and pathology for an oncology case).
  • Shows comparison scorecard [15:14] evaluated via an Opus judge agent across details matched, detail match score, personal voice score, and insightfulness.
• Conclusion & Summary [17:47 - 18:51]: Final recommendations breakdown by workflow type.

Claims & numbers

• The presenter states that with the release of Gemini 4 and Muse Wonder, the frontier AI race is now a “four-horse race” [00:27].
• The presenter notes that Meta’s Chief AI Officer Alex Wang stated that Muse Wonder was not trained on 3D models [03:22].
• In the writing benchmark scored out of 31 interview transcript details [15:14]:
  • Gemini 4: Matched 21/31 details, detail match score 7/10, personal score 7/10, insightful score 8/10, overall 6.7/10.
  • Opus 5.5: Matched 22/31 details, detail match score 7/10, personal score 4/10, insightful score 9/10, overall 6.3/10.
  • GPT-6 Astra: Matched 20/31 details, detail match score 6/10, personal score 5/10, insightful score 7/10, overall 5.7/10.
• The presenter claims Gemini 4’s writing is significantly improved and far less verbose compared to Gemini 3.8 [16:16].

Notable quotes

• [00:17] “The frontier model race is no longer just between OpenAI and Anthropic. With today’s release of Gemini 4 and the launch of Muse Wonder, it’s now a four-horse race.”
• [03:39] “In my humble opinion, Opus is the clear winner here, with Gemini as the runner-up, Astra in number three, and Muse as number four.”
• [17:49] “Opus is still best-in-class on 3D models, animations, and, you know, editing videos and adding music... Opus is probably the most flashy right now.”

Assessment

This is an authentic hands-on product review and benchmark comparison video by creator Peter Yang testing real outputs generated by Claude Opus 5.5, GPT-6 Astra, Gemini 4, and Muse Wonder. The code-rendered Three.js demos, Blender cinemats, and HTML websites are actively interacted with and demonstrated live on screen.

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