Gemini 4 Tops Writing Test as Claude Opus 5.5 Sweeps 3D Generation

Peter Yang

Anthropic's Claude Opus 5.5 retains technical dominance in procedural 3D coding and game design, but Google's Gemini 4 unseats rivals in everyday workflows by capturing natural writing voice and complex research synthesis.

AI-assisted summary of I Tested Gemini 4 vs Opus 5.5 vs GPT-6 Astra vs Muse Wonder (5 Use Cases) — original source. Check the original for full context.

Source published: 2026-10-13

Key ideas

The frontier AI benchmark landscape has officially expanded into a four-horse race between Claude Opus 5.5, Gemini 4, GPT-6 Astra, and Muse Wonder, with no single model leading across every multimodal category. (0:00)
Claude Opus 5.5 swept procedural 3D modeling and game development, delivering superior one-shot Blender environments and a playable Three.js rail shooter with custom audio and boss mechanics. (0:43)
Gemini 4 claimed first place in human-like long-form newsletter writing and complex biomedical research, overcoming prior verbosity issues to produce natural voice and exhaustive technical synthesis. (14:16)
GPT-6 Astra captured the web design benchmark through integrated image generation and minimalist layout, while Muse Wonder took the animated storytelling round with a poignant narrative twist despite primitive 3D capabilities. (11:20)

Detailed notes

## Background Published on YouTube by Peter Yang on October 13, 2026, this video examines the newly expanded four-horse frontier AI race following the release of Google's Gemini 4 and Meta's Muse Wonder. According to the page description, Yang tests Gemini 4, Claude Opus 5.5, GPT-6 Astra, and Muse Wonder across five real-world creative and technical domains: 3D Blender modeling, playable game development, coded animated shorts, website design, and long-form writing.

## TL;DR Peter Yang systematically evaluates four frontier AI models across five complex code-generation and creative benchmarks to identify where each system actually excels. Claude Opus 5.5 leads the pack in raw procedural multimedia generation, taking first place in both the 3D Himeji Castle simulation and the interactive Star Fox-inspired rail shooter game. GPT-6 Astra captures the website redesign category by leveraging clean layouts and integrated multi-modal imagery, while Muse Wonder delivers a surprise win in narrative emotional impact despite primitive rendering. Gemini 4 proves to be the standout surprise of the suite, claiming first place in long-form newsletter drafting by capturing an authentic human voice and demonstrating unmatched depth in high-stakes medical research synthesis.

## Atmosphere & style The presentation carries an energetic, pragmatic, and maker-focused tone, trading marketing hype for direct code execution and side-by-side artifact comparisons. Yang's delivery is candid and analytical, openly highlighting bizarre quirks—like accidental visual bugs and cluttered user interfaces—with lighthearted humor. The technical momentum takes on a sincere, grounded emotional resonance when exploring Muse Wonder's poignant narrative twist and Yang's personal account of turning to Gemini 4 during his mother's medical challenges, making the benchmark feel deeply relevant to real-world workflows.

## Bulletin - [00:00](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=0s) The frontier AI benchmark landscape has officially expanded into a four-horse race between Claude Opus 5.5, Gemini 4, GPT-6 Astra, and Muse Wonder, with no single model leading across every multimodal category. - [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) Claude Opus 5.5 swept procedural 3D modeling and game development, delivering superior one-shot Blender environments and a playable Three.js rail shooter with custom audio and boss mechanics. - [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) Gemini 4 claimed first place in human-like long-form newsletter writing and complex biomedical research, overcoming prior verbosity issues to produce natural voice and exhaustive technical synthesis. - [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) GPT-6 Astra captured the web design benchmark through integrated image generation and minimalist layout, while Muse Wonder took the animated storytelling round with a poignant narrative twist despite primitive 3D capabilities.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=0s) The frontier AI race has diversified beyond Anthropic and OpenAI into a four-way competition that includes Google's Gemini 4 and Meta's Muse Wonder. - [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) Claude Opus 5.5 won the 3D procedural modeling benchmark, creating a multi-layered Himeji Castle render complete with falling snow, cinematic camera sweeps, and coded traditional music. - [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) Meta's Muse Wonder finished last in both 3D Blender modeling and game rendering due to a lack of dedicated 3D mesh training data. - [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) In interactive game generation, Opus 5.5 built a fully playable Star Fox x StarCraft rail shooter featuring custom flight physics, wingman dialog boxes, and end-of-stage boss mechanics. - [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) Gemini 4 exhibited high mechanical complexity in game design, incorporating cockpit camera views and Battlecruiser support strikes, but suffered from an overcrowded visual interface. - [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) Muse Wonder won the coded animated short test by prioritizing emotional narrative pacing and a graveyard reveal over raw graphical fidelity. - [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) GPT-6 Astra's animated short fell into last place due to an unintended visual rendering bug that made a plant-watering robot appear to be urinating. - [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) GPT-6 Astra secured first place in brand and website redesign by pairing clean landing page aesthetics with native AI image generation for guest portraits and graphics. - [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) Gemini 4 tied for third place in website design because its eagerness to include every element resulted in an overwhelming wall of copy. - [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) In a blind evaluation against a human baseline, Gemini 4 took first place in long-form writing by nailing conversational anecdotes, quote selection, and authentic author voice. - [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) Claude Opus 5.5 offered the strongest analytical breakdowns on transcript summarization but scored last on warmth and sounding like a personal voice. - [16:31](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=991s) Gemini 4 demonstrated a notable advantage in high-stakes biomedical research, surfacing specialized diagnostic angles and clinical research nuances that Opus 5.5 and Astra missed. - [17:45](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=1065s) The frontier model landscape has fractured into distinct specializations: Opus 5.5 for procedural multimedia and 3D code, Astra for UI design, and Gemini 4 for authentic writing and dense technical research.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=0s) The New Four-Horse Frontier Race

The frontier AI model landscape is no longer just a two-way duel between OpenAI and Anthropic. With the release of Gemini 4 and Meta's launch of Muse Wonder, the competition has officially expanded into a four-horse race alongside Claude Opus 5.5 and GPT-6 Astra. Having secured early access to all contenders except Muse, Peter Yang evaluates how these four cutting-edge models handle five rigorous, real-world creative and technical challenges: procedural 3D modeling, interactive game development, coded animated shorts, website design, and writing.

```mermaid flowchart TD accTitle: Frontier Model Four-Horse Race Testing Pipeline accDescr: Diagram showing the four competing frontier AI models being evaluated across five distinct multimodal code-generation use cases.

subgraph Models["Frontier AI Models"] M1["Anthropic Opus 5.5"] M2["Google Gemini 4"] M3["OpenAI GPT-6 Astra"] M4["Meta Muse Wonder"] end

subgraph Tests["Five Evaluation Benchmarks"] T1["Test 1: 3D Himeji Castle"] T2["Test 2: Star Fox x StarCraft Game"] T3["Test 3: Emotional Animated Short"] T4["Test 4: Website Brand Redesign"] T5["Test 5: Long-Form Writing"] end

M1 & M2 & M3 & M4 --> T1 M1 & M2 & M3 & M4 --> T2 M1 & M2 & M3 & M4 --> T3 M1 & M2 & M3 & M4 --> T4 M1 & M2 & M3 & M4 --> T5 ```

[00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) Test 1: Himeji Castle in Blender with Procedural Music

To push the limits of spatial understanding and algorithmic artistry, the first benchmark tasks each model with modeling Japan's historic Himeji Castle in Blender and generating a 30-second dawn cinematic during cherry blossom season. The models are instructed to capture the castle's distinctive white plaster walls and complex multi-layered rooflines faithfully, while composing an original accompanying soundtrack—all written and rendered entirely through code without external assets.

Opus 5.5 leads the pack with an exceptional render that feels like an authentic Japanese tourism commercial. Its sweeping camera cuts move gracefully past cherry blossom branches toward the castle keep, set against subtle falling snow and traditional instrumentation. Gemini 4 follows closely behind as the clear runner-up, executing varied and dynamic camera angles—including low shots beneath an arched bridge—paired with traditional Japanese flute compositions, though its visual finish sits just behind Opus.

GPT-6 Astra claims third place. While Astra captures delicate water reflections and ambient birdsong, its castle geometry appears noticeably flat, and its cinematography lacks the cinematic flair demonstrated by Opus and Gemini. Meta's Muse Wonder lands in a distant last place with primitive, muddy visuals reminiscent of models from several generations prior. This poor output aligns with statements from Meta Chief AI Officer Alexandr Wang, who noted that Muse was not specifically trained on 3D mesh generation.

| Model | Test 1: 3D Himeji Castle (Blender & Audio) | Test 2: Star Fox x StarCraft Rail Shooter | Test 3: Emotional Animated Short | |:--- |:--- |:--- |:--- | | **Opus 5.5** | **Winner (1st)**: Authentic tiered architecture, falling snow, sweeping dawn cinematography, traditional flute score. | **Winner (1st)**: Cinematic camera, dynamic wingman chatter, polished boss fight, responsive one-shot build. | **Runner-Up (2nd)**: Highest visual and lighting polish; story about a waiting spouse is sweet but emotionally modest. | | **Gemini 4** | **Runner-Up (2nd)**: Excellent bridge and ground angles, traditional flute; slightly less polished than Opus. | **Runner-Up (2nd)**: Feature-dense (cockpit view, Battlecruiser summon, barrel rolls), crowded UI layout. | **Third (3rd)**: Poetic symbolism with flying doves; focused on work-life estrangement, highly artistic. | | **GPT-6 Astra** | **Third (3rd)**: Realistic water reflections and birdsong, but castle structure looks flat and camera work is basic. | **Third (3rd)**: Atmospheric and moody lighting, but limited to a single enemy type and basic terrain. | **Fourth (4th)**: Charming robot watering a plant; undercut by an accidental visual bug resembling urination. | | **Muse Wonder** | **Fourth (4th)**: Outdated graphics; severe geometric warping due to a lack of dedicated 3D training data. | **Fourth (4th)**: Extremely primitive visuals and rudimentary gameplay, resembling early legacy models. | **Winner (1st)**: Understated art style, but delivers an unexpected, poignant twist involving a memorial tombstone. |

[03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) Test 2: A Star Fox x StarCraft Rail Shooter

The second test raises the stakes by combining two classic video game franchises into an interactive three.js rail shooter coded from scratch. Inspired by the creator's favorite game, *StarCraft*, the prompt requests a *Star Fox*-style arcade dogfight where a Terran Wraith ship defends against swarms of airborne Zerg, complete with custom Blender-modeled ships, barrel rolls, laser projectiles, background terrain, a boss encounter, and a programmatic soundtrack.

Opus 5.5 once again delivers the most cohesive, playable package in a single prompt. The game features engaging flight dynamics, custom sound effects, bottom-left radio chatter from a wingman, and an imposing end-of-stage boss firing energy barrages. While its visual presentation suffers slightly from heavy lens flares and overly reflective materials, it plays smoothly straight out of the gate.

Gemini 4 secures second place through sheer feature density. While its screen layout feels crowded and its visual polish lags behind Opus, Gemini incorporates an impressive suite of one-shot gameplay mechanics. Players can toggle into an immersive first-person cockpit perspective by pressing `V`, perform evasive barrel rolls, and press `B` to call in a Terran Battlecruiser that obliterates enemies with a Yamato Cannon. Gemini acts like a model determined to outwork its rivals to impress the user.

GPT-6 Astra ranks third, capturing moody sci-fi atmospheric lighting but falling short on gameplay variety with only one basic Zerg flyer and barren landscapes. Meta's Muse Wonder finishes last, generating a barebones, clunky demo that feels like an artifact from an early predecessor model like Claude 3.1 Opus.

- **Opus 5.5 highlights**: One-shot boss encounter, dynamic wingman dialog UI, cinematic flight camera, responsive laser firing (marred only by excess lens flare). - **Gemini 4 features**: First-person cockpit mode (`V` key), tactical Battlecruiser support strikes (`B` key), high mechanical complexity. - **GPT-6 Astra features**: Strong atmospheric tone, but constrained by minimal enemy variety and flat terrain geometry. - **Muse Wonder features**: Extremely rudimentary three.js rendering, highlighting Meta's lack of 3D gaming optimizations.

[07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) Test 3: An Animated Emotional Short

The third benchmark evaluates narrative depth and emotional resonance by instructing each model to write, animate, and score a 30-second short film entirely via code, strictly avoiding pre-rendered or stock assets.

Opus 5.5 showcases the highest technical production value. Its animation portrays a woman waiting at home late into the night, keeping a lit lantern burning for her absent husband. While visually pristine and evocative in its lighting, the narrative arc feels conventional, stirring mild sympathy rather than genuine heartache. Gemini 4 pivots toward poetic abstraction, animating soaring doves and changing vistas to illustrate a person consumed by professional toil who slowly grows estranged from their loved ones. Its art direction is striking, but the emotional core remains somewhat muted.

GPT-6 Astra takes a narrative detour with an animated short about a small robot dutifully watering a sapling until the plant towers overhead. While Astra attempts a bittersweet story about mechanical self-sacrifice, the visual execution creates an unintended comedic moment where the robot appears to be "peeing into a plant" rather than irrigating it. This visual bug, combined with an abrupt revival sequence at the end, undermines the dramatic tension.

The unexpected victor of the round is Meta's Muse Wonder. While its raw graphics and shader complexity fall well short of Opus, its storytelling delivers an authentic emotional gut-punch. The short initially follows an elderly man tenderly cultivating a solitary flower. In the final seconds, the camera pulls back to reveal that he has grown the blossom solely to lay it upon the gravestone of his deceased spouse. The sudden reveal delivers a poignant narrative payoff, proving that thoughtful pacing and emotional storytelling can triumph over raw graphical horsepower.

[11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) Test 4: Brand & Website Redesign for Behind the Craft

Moving from visual arts to practical digital product design, the fourth test challenges the models to execute a full brand identity and homepage redesign for *Behind the Craft*—Peter Yang's newsletter, YouTube channel, and podcast. The prompt requires a fresh logo, a unified color palette and typography system, and an original homepage layout that avoids generic AI design tropes, all created without looking at the existing live website.

Opus 5.5 produces an intriguing conceptual layout with an interactive peel-back hero section: users can drag a physical page corner to reveal the underlying tagline, "Learn AI by watching it built." Beyond this clever introductory flourish, however, the rest of the page settles into a conventional, somewhat uninspired structure. Peter rates Opus's submission as a "B effort," noting that Claude's lack of a native image generation pipeline leaves the layout feeling text-heavy and devoid of supporting graphical assets.

When evaluating Gemini 4's take on the redesign, the model immediately struggles with layout density, burying the value proposition under an overwhelming wall of copy right out of the gate.

[12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) Test 4 Verdict: Website Redesign for Behind the Craft

When reviewing Gemini 4's website redesign for *Behind the Craft*, the immediate impression is one of visual clutter. While Gemini successfully captured every required element—including Peter's courses, a complete design system, specific typography, and custom color palettes—the homepage ended up feeling completely overstuffed with copy. It mirrored the rail shooter game Gemini built earlier in the tests: an eagerness to generate a massive volume of features that ultimately overwhelmed the user. Gemini proved to be the model that "tries the hardest" out of all four, persistently attempting to satisfy the user, but in this specific instance it simply went overboard.

GPT-6 Astra took the exact opposite approach and delivered a standout layout. Astra produced a clean, disciplined landing page centered around the tagline "Make something of AI," paired with a prominent email capture call-to-action. What set Astra apart was its integration of ChatGPT image generation to populate the page with believable illustrations and portraits of past interview guests. While OpenAI models have historically lacked refined aesthetic taste for web design, Astra's output felt polished and uncluttered.

Muse Wonder sat on the other end of the spectrum, producing a design that looked entirely like dense, unformatted newspaper text rather than a modern creator landing page. With Astra capturing the crown, the final standings for web design placed Astra in first, Claude Opus 5.5 in second, and Gemini 4 tied with Muse Wonder in third place.

---

[14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) Test 5: Long-Form Writing and Newsletter Extraction

Moving away from visual assets into everyday workflows, the fifth test evaluated how well each frontier model handled core writing tasks. Peter ran a production prompt designed to turn a raw, unstructured interview transcript (from a conversation with Sue) into a finished newsletter post consisting of an introductory framing narrative and a breakdown of core takeaway insights. Historically, achieving a polished one-shot draft without tedious human back-and-forth has been nearly impossible, so the test measured which model could come closest to Peter's manually edited final publication.

To score the drafts objectively, Peter had an isolated Claude Opus sub-agent—operating completely blind with zero conversational memory—grade the drafts from Gemini 4, Claude Opus 5.5, and GPT-6 Astra against the human baseline across three criteria: personal voice, insightfulness, and detail fidelity. Surprisingly, Gemini 4 took first place over Opus and Astra. The evaluator noted that Gemini produced "the only draft that sounds like you," successfully selecting the right verbatim opening quote, incorporating personal anecdotes, and matching the rhythm of the newsletter, even if a few traces of jargon lingered.

Opus 5.5 scored highest on sheer analytical insight but ranked dead last on sounding personal or relatable. GPT-6 Astra produced a factually accurate summary but fell flat on both insight and personal resonance. This victory highlighted a major generational jump for Google: while Gemini 3.8 suffered from notorious verbosity and repetitive phrasing, the Google team's focused tuning on writing voice made Gemini 4 significantly sharper and more natural.

- **Gemini 4**: Matched personal voice, selected authentic opening quotes, nailed conversational anecdotes; took 1st place. - **Claude Opus 5.5**: Delivered the deepest thematic insights, but read as overly detached and cold. - **GPT-6 Astra**: Retained raw transcript accuracy, but lacked analytical depth and authentic personality.

---

[16:31](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=991s) High-Stakes Research: Gemini 4's Medical Breakthrough

Gemini 4's updated writing engine carries over directly into high-stakes research workflows. Peter shared a personal development: his mother had been dealing with ongoing health complications, prompting him to run intensive medical queries across frontier AI models to identify potential treatment pathways and diagnostic angles.

When pitted against the other models on these technical inquiries, Gemini 4 delivered an unexpectedly superior performance. It surfaced specific diagnostic paths, clinical research nuances, and medical details that Opus 5.5 and Astra completely overlooked.

Historically, Gemini's biggest liability was its extreme verbosity and tendency to output sprawling blocks of text. However, for dense biomedical synthesis and complex literature analysis, that natural thoroughness transforms into an advantage. When combined with Gemini 4's upgraded writing clarity, that comprehensiveness makes it a standout tool for heavy research synthesis and technical writing.

---

[17:45](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=1065s) Final Verdict: The Four-Horse Frontier Race Decided

With five demanding tests completed across procedural 3D modeling, interactive game development, coded animation, web design, and writing, the frontier model landscape is no longer dominated by a single clear leader. Each engine has established its own distinct competitive domain based on practical utility:

```mermaid flowchart TD accTitle: Frontier AI Model Specializations accDescr: Diagram showing the top practical use cases for Opus 5.5, GPT-6 Astra, Gemini 4, and Muse Wonder.

subgraph Frontier["Frontier Model Sweet Spots"] M1["Claude Opus 5.5"] --> F1["3D Procedural Assets, Blender Scripting, Audio & Video"] M2["GPT-6 Astra"] --> F2["Clean Web UI & Integrated Multi-modal Generation"] M3["Gemini 4"] --> F3["Natural Voice Writing & Deep Medical/Technical Research"] M4["Muse Wonder"] --> F4["Autonomous Personal Agent Workflows (Muse App)"] end ```

Claude Opus 5.5 remains best-in-class for creative, procedural multimedia—dominating Three.js scenes, Blender geometry, soundtrack composition, and cinematic animations. GPT-6 Astra claimed the design crown due to its balanced UI instincts and ChatGPT image integration, although Peter noted Google could rival this if they properly pipe tools like Nano Banana into Gemini's web generation flows.

Gemini 4 emerged as the preferred engine for day-to-day knowledge work, taking the top spot in both human-sounding newsletter writing and deep research. Meanwhile, Muse Wonder's real strengths appear tied to its personal agent features inside the dedicated Muse ecosystem, which Peter plans to evaluate in a separate deep dive.

## Notable quotes > [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) 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. — Peter Yang

> [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) 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. — Peter Yang

> [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) Switch up the level to Terrans against the Zerg, and model all the Zerg properly in Blender. The Terran ship can be a Wraith. — Peter Yang

> [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) 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. — Peter Yang

> [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) peeing into a plant — Peter Yang

> [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) 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. — Peter Yang

> [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) Learn AI by watching it built.

> [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) B effort — Peter Yang

> [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) tries the hardest — Peter Yang

> [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) Make something of AI

> [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) the only draft that sounds like you — Claude Opus sub-agent

## Mentioned Peter Yang, Alexandr Wang, Sue, Anthropic, Google, Meta, OpenAI, Gemini 4, Claude Opus 5.5, GPT-6 Astra, Muse Wonder, Claude 3.1 Opus, Gemini 3.8, Blender, three.js, ChatGPT, Nano Banana, Behind the Craft, Star Fox, StarCraft, Himeji Castle

## Video description Copy my best AI workflows to save time and automate busywork: https://www.behindthecraft.com 🤝 Subscribe to my practical AI newsletter read by 150K+ professionals: https://creatoreconomy.so

Gemini 4 just launched and the frontier model race is now a four-horse race. I tested Gemini 4, Opus 5.5, GPT-6 Astra, and Muse Wonder across five real use cases: 3D models, games, animated shorts, website design, and writing.

What I cover: (00:00) Four frontier models build Himeji Castle in code (00:43) Test 1: Himeji Castle in Blender with music (03:46) Test 2: A Star Fox x StarCraft rail shooter (07:02) Test 3: An animated short that makes you feel something (11:20) Test 4: Which model is best at website design (14:16) Test 5: Testing each model's writing capabilities (16:31) The use case that really surprised me about Gemini 4 (17:45) Recapping the winners for each use case

Prompts I used:

Himeji Castle: "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."

Star Fox x StarCraft: "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."

Emotional short: "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."

Website redesign: "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."

Where to find me: X: https://x.com/petergyang LinkedIn: https://www.linkedin.com/in/petergyang/

Subscribe to this channel - more tutorials coming soon!

[Source](https://www.youtube.com/watch?v=zueuzhBkd9Y)

中文简报

Gemini 4领跑写作测试，Claude Opus 5.5横扫3D生成 Anthropic的Claude Opus 5.5在程序化3D编码和游戏设计方面保持技术优势，而Google的Gemini 4则凭借自然的写作语气和复杂的综合研究能力，在日常工作流程中超越竞争对手。 Peter Yang在五个复杂的代码生成与创意基准测试中对四款前沿AI模型进行了系统评估，以明确各系统真正的优势所在。Claude Opus 5.5在纯程序化多媒体生成方面领跑，在3D姬路城模拟和受Star Fox启发的交互式轨道射击游戏中均拔得头筹。GPT-6 Astra凭借简洁的布局和集成的多模态图像在网站重新设计项目中夺冠，而Muse Wonder虽然渲染效果简陋，却在叙事情感冲击力上意外获胜。Gemini 4成为测评中最令人惊喜的黑马，它以高度逼真自然的人类语气撰写长篇时事通讯，并在高难度医学研究综合分析中展现出无可比拟的深度，成功位列第一。前沿AI基准测试格局已正式演变为Claude Opus 5.5、Gemini 4、GPT-6 Astra和Muse Wonder之间的“四方争霸”，目前没有任何单一模型能在所有多模态类别中全面领先。 (0:00) Claude Opus 5.5横扫程序化3D建模与游戏开发领域，不仅实现了优异的单次生成Blender环境，还构建了一款包含自定义音频及Boss战机制的可玩Three.js轨道射击游戏。 (0:43) Gemini 4在拟人化长篇时事通讯撰写和复杂生物医学研究方面摘得桂冠，克服了以往过于冗长的问题，展现出自然的语言风格和详尽的技术综合能力。 (14:16) GPT-6 Astra通过集成图像生成和极简布局赢得了网页设计基准测试，而Muse Wonder尽管3D能力较为原始，却凭借感人至深的叙事转折拿下了动画叙事测试的胜利。 (11:20)

Reading summary

Testing across five technical and creative benchmarks indicates that the frontier AI race has split into specialized domains rather than yielding a single dominant model, according to evaluations by Peter Yang. Procedural 3D and Interactive Game Engines (0:43) In benchmarks requiring pure code generation without external assets, Claude Opus 5.5 consistently outperformed rivals in spatial reasoning and multimedia cohesion. For a 30-second Blender dawn cinematic of Himeji Castle featuring programmatic music, Opus generated authentic tiered architecture, animated falling snow, and a traditional score. It repeated this performance in a three.js rail shooter combining elements of Star Fox and StarCraft, outputting responsive flight controls, wingman radio chatter, and a multi-phase boss fight from a single prompt. Gemini 4 secured second place in both tests through sheer mechanical density, adding dynamic camera sweeps to the castle render and integrating cockpit views and Battlecruiser support strikes into the game demo. However, its visual presentation suffered from interface clutter. GPT-6 Astra ranked third, constrained by flat castle geometry and minimal enemy variety. Meta's Muse Wonder placed last in both procedural 3D tests due to severe geometric warping and crude assets, aligning with statements from Meta Chief AI Officer Alexandr Wang that the model lacked dedicated 3D mesh training. Narrative Resonance Over Raw Graphical Polish (7:02) Evaluating narrative depth through code-animated 30-second short films highlighted a clear divergence between graphical fidelity and emotional storytelling. Opus 5.5 delivered the highest technical polish and lighting in a story about a woman waiting for her absent husband, but the plot remained conventional. Gemini 4 opted for symbolic abstraction depicting workplace estrangement with flying doves, while GPT-6 Astra's story about a robot watering a plant was undercut by an unintended visual glitch resembling the character urinating. Muse Wonder unexpectedly won the animation benchmark despite exhibiting primitive graphics and shader limitations. The short followed an elderly man cultivating a single flower, culminating in a camera pull-back that revealed he was placing it on his deceased spouse's gravestone. The twist demonstrated that effective narrative pacing and emotional framing can outweigh raw graphical horsepower. Web Design and Layout Discipline (12:26) When tasked with creating a full brand identity and homepage for Peter Yang's media property Behind the Craft, GPT-6 Astra claimed first place. Astra produced a clean, disciplined layout centered around the tagline 'Make something of AI' and a clear email call-to-action. Its key advantage was native ChatGPT image generation, which populated the site with coherent illustrations and guest portraits. Opus 5.5 placed second; while it featured an interactive peel-back hero section, the absence of native image generation left the page looking text-heavy. Gemini 4 and Muse Wonder tied for third place. Although Gemini successfully generated an entire design system and course catalog, it overwhelmed the interface with excessive walls of text. Muse Wonder finished with a flat, unformatted layout that resembled a newspaper article rather than an interactive website. Long-Form Writing Voice and Complex Research (14:16) In text generation workflows, Gemini 4 showed significant generational progress over its predecessor's verbosity. When transforming an unstructured interview transcript into a structured newsletter post, an isolated, blind Claude Opus evaluation sub-agent ranked Gemini 4 ahead of Opus 5.5 and Astra. The evaluator noted that Gemini was the only model to capture the author's conversational voice, cadence, and personal anecdotes, whereas Opus 5.5 felt cold and analytical, and Astra lacked narrative depth. This synthesis capability transferred directly to high-stakes technical inquiries. When tested on complex diagnostic questions regarding health complications faced by Yang's mother, Gemini 4 surfaced specific clinical research nuances and diagnostic angles that both Opus 5.5 and Astra overlooked. For dense biomedical synthesis, Gemini's thoroughness proved to be a practical advantage. Yang concludes that frontier model selection now depends on distinct use cases rather than overall benchmark dominance: Opus 5.5 leads in procedural 3D modeling and game generation; Astra excels in clean UI layout and multimodal design; and Gemini 4 serves as the strongest engine for voice-matched writing and deep research. Muse Wonder's broader utility remains tied to its dedicated personal agent ecosystem, which was not evaluated in these standalone code benchmarks.

Detailed notes — zh

## Background 该内容由 Peter Yang 于 2026 年 10 月 13 日发布在 YouTube 频道。视频简介显示，随着谷歌推出 Gemini 4，前沿大模型竞赛进入“四雄争霸”阶段，作者对 Gemini 4、Claude Opus 5.5、GPT-6 Astra 和 Meta 的 Muse Wonder 进行了横向对比实测，涵盖 3D 建模、游戏构建、代码动画短片、网站设计与文案写作五大实际用例。 ## TL;DR 本次评测通过五个高难度真实任务全面考察了四大前沿模型的工程实现与生成能力，结果显示各模型专长分化显著，并无绝对的全能霸主。Claude Opus 5.5 在 Blender 3D 场景渲染和 Three.js 轨道射击游戏开发中展现出顶级的视觉审美与代码工程实力；GPT-6 Astra 凭借干净的极简排版与原生图像生成在网站改版任务中胜出。Meta 的 Muse Wonder 虽在 3D 和代码层面整体落后，却在 30 秒情感动画测试中凭借出人意料的叙事反转爆冷夺魁；而 Gemini 4 则在文风拟人化的长文写作以及深度医学调研中表现出惊人的严密性与统治力。 ## Atmosphere & style 整个评测呈现出硬核极客、客观实操且充满探索趣味的节奏氛围。视频中既有目睹 Opus 5.5 与 Gemini 4 凭纯代码生成电影级古建筑与复杂战机座舱机制时的惊叹，也有看到 Astra 机器人浇花建模失误引发啼笑皆非的幽默时刻，更有被 Muse Wonder 意外催泪的反转短片击中情感的戏剧性张力。评测者以快节奏、交付物说话的方式推进，让观众清晰且直观地体会到各大模型如同不同性格的工程师或创作者，正处于能力狂飙突进的技术拐点之中。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=0s) 前沿大模型竞争从 OpenAI 与 Anthropic 的双雄争霸转为包含 Gemini 4 与 Muse Wonder 在内的四方角逐，各模型在 3D、代码、设计和写作等维度呈现出显著的专长分化。 - [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) Claude Opus 5.5 在 Blender 3D 建模与全代码小游戏开发测试中拔得头筹，展现出极高的物理结构还原度与完备的工程代码交付能力。 - [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) Gemini 4 在长文写作盲测中击败 Opus 5.5 与 GPT-6 Astra，并在复杂医学调研场景中挖掘出其他模型遗漏的关键信息，成为深度研究与文案的核心工具。 - [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) GPT-6 Astra 凭借简洁的现代排版与原生图像生成结合能力斩获网站设计测试第一名，而 Meta Muse Wonder 则在情感叙事动画短片测试中爆冷夺魁。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=0s) 大模型竞争格局正式演变为 Opus 5.5、GPT-6 Astra、Gemini 4 和 Muse Wonder 四强并立，单一模型不再全面统治所有任务场景。 - [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) 在 Blender 姬路城 3D 渲染与配乐测试中，Opus 5.5 凭借电影级画面质感与准确的日式古建筑形制夺得第一名。 - [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) Gemini 4 在 3D 运镜调度上展现出丰富的镜头语言，并能自动使用代码编写契合氛围的日式尺八竹笛配乐。 - [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) Meta Muse Wonder 在 3D 建模方面表现落后，因其底层未针对此类 3D 几何渲染进行专项训练。 - [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) Opus 5.5 具备卓越的单次提示（One-shot）代码生成能力，能够完整交付包含 Boss 战循环与僚机通信弹窗的轨道射击游戏。 - [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) Gemini 4 在代码构建中展现出极强的主动性与功能堆料倾向，塞入了第一人称座舱切换及呼叫战列巡洋舰轰炸等复杂机制。 - [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) 在 30 秒全代码生成动画短片测试中，Muse Wonder 凭借老人为亡妻墓碑献花的情感反转剧情力压画质更优的对手，爆冷夺冠。 - [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) GPT-6 Astra 在动画短片生成中出现了机器人浇水动作疑似小便的严重建模 Bug，削弱了原本的悲情叙事表达。 - [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) GPT-6 Astra 在品牌主页改版测试中以极简明快的布局、精准的号召性用语以及原生嘉宾头像插画生成摘得设计第一名。 - [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) Gemini 4 在网页设计中存在文字信息过载、缺乏留白的问题，与 Muse Wonder 报纸般的纯文本排版一同在设计项中垫底。 - [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) 在无上下文的访谈转 Newsletter 盲测中，Opus 子代理评定 Gemini 4 的文风最贴近人类作者原声，打破了以往版本假大空与啰嗦的固有印象。 - [16:31](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=991s) Gemini 4 在复杂医学文献检索与方案推导中展现出统治级能力，其长篇详尽的生成风格在严谨调研场景下转化为独特优势。 - [17:45](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=1065s) 当前大模型选型建议依据具体场景划分：追求 3D 与视听多媒体资产选 Opus 5.5，界面设计与落地页选 GPT-6 Astra，专业研究与深度长文写作首选 Gemini 4。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=0s) 四雄争霸：前沿大模型迎来全新竞争格局随着谷歌最新发布 Gemini 4 以及 Meta 推出 Muse Wonder，AI 前沿模型的竞技场彻底告别了过去由 OpenAI 与 Anthropic 两强争霸的局面，正式演变成一场激烈的“四马奔腾”（four-horse race）。除了 Meta 的 Muse Wonder 之外，测评者获得了其余三款模型的抢先体验资格。为了全面考察它们的真实实力，本次测评设计了涵盖五个维度的硬核实测：3D 建模、游戏构建、代码生成动画短片、网站设计以及文案写作。测试的起点是一个极具挑战性的视觉与代码生成任务——通过纯代码方式在 Blender 中还原日本著名的姬路城（Himeji Castle）。参赛选手包括 Anthropic 的 Claude Opus 5.5、OpenAI 的 GPT-6 Astra、谷歌的 Gemini 4 以及 Meta 的 Muse Wonder。这场测试不仅比拼模型的几何构建与材质渲染能力，还要考察其镜头语言与由代码编排的配乐水平。 --- ### [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) 核心测试一：Blender 中的日本姬路城 3D 渲染与配乐姬路城是日本最具代表性的古建筑之一，其经典的白漆外墙与复杂的层叠飞檐对 3D 建模而言极其严苛，极其考验模型对真实物理结构和细节比例的理解。给四款模型的提示词统一为：“在 Blender 中构建姬路城，并渲染一段 30 秒的清晨樱花季电影感短片。确保建筑形制准确，并完全使用代码生成配乐。” 在实际渲染展示中，Opus 5.5 拿出了堪称全场最佳的震撼表现。画面从掩映的樱花树枝轻柔推向姬路城主体，天空中还飘落着细腻的花瓣与雪花，建筑本体的还原度极其惊人，完美再现了现实中的风貌，整体质感“简直就像一部正规的日本旅游宣传大片”。紧随其后的 Gemini 4 表现同样可圈可点，不仅精准把握了城堡的外观，还给出了更为丰富的运镜调度，例如穿过桥底的仰拍视角，且与 Opus 5.5 不约而同地生成了韵味悠长的日本传统尺八竹笛配乐。相比之下，GPT-6 Astra 虽在水面倒影和清脆鸟鸣的环境音效上可圈可点，但姬路城的主体建模略显扁平，失去了真实城堡的立体雄伟感，镜头调度也显得相对平淡。垫底的则是 Meta 的 Muse Wonder，其渲染画面如同“几代之前的落后模型产物”，模型粗糙且结构严重失真。虽然 Meta 首席 AI 官 Alex Wang 曾公开表示 Muse 并没有针对 3D 建模进行专项训练，但这一轮测试中它与其他三者的差距依然十分悬殊。 | 模型 | 3D 建模准确度 | 运镜与画面质感 | 代码生成音频表现 | 本轮排名 | |:--- |:--- |:--- |:--- |:--- | | **Opus 5.5** | 极高（精准复刻白墙飞檐） | 电影级质感，樱花与雪花层次丰富 | 优雅的传统日本笛乐 | 第 1 名（优胜） | | **Gemini 4** | 很高（造型准确真实） | 镜头视角丰富，包含桥底仰拍 | 传统日式风情配乐 | 第 2 名（次席） | | **GPT-6 Astra** | 一般（建筑偏平，缺乏立体细节） | 视角偏保守，但水面倒影出色 | 带有环境鸟鸣与悠扬音乐 | 第 3 名 | | **Muse Wonder** | 极差（宛如多年前的老旧模型） | 粗糙生硬，缺乏空间与光影层次 | 基础音频生成 | 第 4 名（垫底） | --- ### [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) 核心测试二：《星际争霸》遇上《星际火狐》——轨道射击游戏实测构建经典轨道射击游戏《星际火狐》（Star Fox）是测评者测试新模型的保留曲目。这一次测试提升了难度，将《星际火狐》的玩法与测评者最爱的经典即时战略游戏《星际争霸》（StarCraft）结合：要求模型使用 Three.js 构建游戏逻辑，并在 Blender 中制作全套资产，场景设定为人族（Terran）驾驶幽灵战机（Wraith）对抗漫天飞舞的虫族（Zerg）异龙与飞龙，包含翻滚动作（Barrel Roll）、激光射击、Boss 战以及独立配乐。 ```mermaid flowchart TD accTitle: Star Fox x StarCraft 游戏测试流程架构 accDescr: 展示四款大模型在结合星际争霸资产与星际火狐轨道射击玩法的代码生成流水线与功能对比。 subgraph Prompt["提示词输入"] P1["设定: 人族战机对抗虫族舰队"] P2["要求: Three.js 逻辑 + Blender 建模"] P3["玩法: 翻滚动作 / Boss 战 / 动态射击 / 原生配乐"] end subgraph Models["四款模型表现"] M_Opus["Opus 5.5 极高完成度 / 镜头声效俱佳 / 僚机对话框"] M_Gemini["Gemini 4 第一人称切换 / 召唤大和巡洋舰 / 功能繁多"] M_Astra["GPT-6 Astra 氛围感强但敌人单一 / 机制简陋"] M_Muse["Muse Wonder 画面简陋 / 机制缺失 / 无法游玩"] end Prompt --> Models ``` 测试中，Opus 5.5 再次凭借强大的综合代码工程能力拔得头筹。它单次提示（One-shot）就交付了一个可玩度极高的关卡：不仅运镜平滑、音效与音乐搭配得当，画面左下方甚至贴心地做出了与僚机通信的实时对话头像框。唯一的瑕疵在于画面反光与镜头光晕（Lens flare）稍微有些过重。 Gemini 4 展现出了截然不同的性格——它像是一个“极力想要给老板留下深刻印象的勤奋员工”，在代码中塞入了令人瞠目结舌的丰富交互机制。玩家不仅能驾驶战机做出翻滚，按“V”键还能无缝切入沉浸感极强的第一人称座舱视角，按“B”键甚至能呼叫配备大和炮的人族战列巡洋舰（Battlecruiser）入场轰炸。虽然界面略显拥挤、抛光度不及 Opus，但其丰富的功能让它稳稳坐上第二把交椅。GPT-6 Astra 虽营造出了不错的战斗暗黑氛围，但机制简陋且敌人种类单一；而 Muse Wonder 生成的游戏再次崩盘，水准甚至退步到了“宛如 Opus 3.1 时代的作品”。 - **Opus 5.5 的高光功能**：完整的关卡流程设计、极富动态感的 Boss 战射击循环、屏幕左下角的动态僚机通讯弹窗。 - **Gemini 4 的特殊机制**：内置按键互动系统（按 V 切换驾驶舱第一人称视角，按 B 召唤战列巡洋舰释放大和炮），功能极度堆料。 - **GPT-6 Astra 的短板**：背景单调，同屏敌人只有一种，游戏玩法深度严重不足。 - **Muse Wonder 的局限**：模型完全未针对此类 3D 渲染及游戏物理逻辑做对齐训练，几乎无法运行。 --- ### [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) 核心测试三：全代码生成感动人心的 30 秒动画短片第三项测试聚焦于大模型的叙事张力与情感表达能力，要求所有模型在不使用任何现成外部素材的前提下，完全通过代码渲染每一帧画面并生成全套音效，制作一部 30 秒、能让观众产生强烈情绪波动的动画短片。四款模型交出了风格迥异的答卷： - **Opus 5.5**：在技术与画质上依然维持了极高水准。短片讲述了一位妻子整夜为未归的丈夫留着一盏明灯的守候故事。画面虽美，但在情感冲击力上略显克制，并没有达到让人瞬间动容的深度。 - **Gemini 4**：走了一条极具艺术审美的视觉路线。短片充斥着飞舞的白鸽与充满隐喻的场景转换，探讨一个为了工作日夜操劳而逐渐疏远挚爱的深刻命题。整体调性极具艺术感，不过在纯粹的情绪煽动力上稍弱于预期。 - **GPT-6 Astra**：设计了一个十分软萌的小机器人悉心浇灌幼苗的故事，主题围绕牺牲与植物的成长展开。然而中间出现了一幕令人啼笑皆非的建模 Bug——机器人的浇水姿势和管道位置“看起来简直就像是在对着植物小便”，严肃悲壮的气氛瞬间变成了滑稽喜剧，削弱了故事本该传达的悲情色彩。 - **Muse Wonder（全场最大反转）**：在前两轮惨败的 Meta Muse Wonder 在本轮爆出惊天大冷门。它讲述了一位白发苍苍的老人日复一日、极其用心地培育一盆植物的故事。短片最后一秒镜头移开，观众赫然发现老人历经千辛万苦养育的花朵，最终被他颤巍巍地献在了亡妻的墓碑前。这种出人意料的情节反转与强烈的情感张力，瞬间打动了测评者，使其力压画质更优的对手，一举夺得本轮最佳。测评者由衷赞叹道，尽管 Muse Wonder 在纯图形与渲染能力上落后于其他对手，但单论讲故事的共情能力与叙事技巧，它绝对是当之无愧的第一名。 --- ### [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) 核心测试四：品牌与网站改版——哪款模型更具设计品味？进入第四个测试用例，测评者要求四大模型在完全不给现有页面作为上下文参考的情况下，为旗下涵盖 Newsletter、YouTube 频道以及播客的个人品牌“Behind the Craft”重新设计整套视觉系统与网站首页。提示词明确要求：“从零开始创建全新的品牌 Logo、配色方案与排版系统，打造一个彻底摆脱‘AI 生成感’（doesn't look AI-generated）的高级首页，且绝对不能参考现有网站。” Opus 5.5 率先登场。它在交互体验上设计了一个非常别致的微动效——页面右上角提示用户“拉开角落以探索幕后”，拖动后会揭晓品牌的核心主旨“Learn AI by watching it built”（通过观看构建过程学习 AI）。这种交互范式极具巧思，但除此亮点之外，整体页面其余排版显得相对保守平淡，测评者给出了“B 级表现”的客观评价。受限于 Claude 自身无法直接原生生成图片的机制，页面在视觉丰富度上也略显单薄。随后登场的是 Gemini 4，然而它在首页排版上的表现却令人大跌眼镜。一打开页面，扑面而来的是极其冗长、密密麻麻的文案大段堆砌，完全失去了现代网页设计应有的留白与呼吸感，整体视觉呈现出严重的文字过载状态。 ### [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) 测试四后续：Gemini 4、Astra 与 Muse 的网页设计横评在评估针对个人品牌“Behind the Craft”的主页重新设计时，Gemini 4 的方案显得有些内容过载。虽然它非常用心地抓取了创作者的课程体系以及所有背景细节，甚至规范化地输出了一套包含配色方案与排版规则（Typography）的设计系统，但页面塞入了过多文案与阅读元素，整体给人的感觉极其拥挤。这与它此前在制作小游戏时的表现如出一辙——试图塞入海量功能与细节，“在四款模型中它绝对是态度最积极、最想让你满意的那个”，但这种过度努力反而让界面显得臃肿杂乱。相比之下，GPT-6 Astra 的设计方案令人眼前一亮。页面走极简明快的路线，提炼出了极具吸引力的标语“Make something of AI”，并配有指向明确的邮箱订阅行动号召（Call to Action）。更巧妙的是，Astra 直接调用了 ChatGPT 的图像生成能力，为过往访谈过的嘉宾绘制了头像与插画配图。尽管 OpenAI 旗下的模型以往并不以视觉设计感见长，但 Astra 呈现出的极简与干净程度在四者中脱颖而出。最后检视 Muse Wonder 的成果，它的表现则相当令人失望。整个页面充斥着密密麻麻的纯文本，排版枯燥乏味，更像是在读一份毫无生气的黑白报纸，完全失去了现代落地页（Landing Page）应有的视觉层次与转化吸引力。 - **网页设计测试最终排名**： - **第一名：GPT-6 Astra**（界面干净简洁，巧妙集成插画与嘉宾头像生成，转化导向明确） - **第二名：Claude Opus 5.5**（排版成熟稳定） - **并列第三：Gemini 4 与 Muse Wonder**（Gemini 堆砌内容严重过载；Muse 全文报纸化、缺乏视觉设计） --- ### [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) 测试五：长篇写作与访谈提炼能力对比告别了炫酷的 3D 建模与视觉生成，评测转向了知识工作者最为高频、朴实的核心场景——文本写作。测试任务是将一段未经剪辑的原始播客访谈速记（Peter 与嘉宾 Sue 的对话记录）转化为标准的 Newsletter 专题长文，要求包含引言概述以及精炼提炼的独家洞察（Takeaways）。在日常工作中，这类任务往往极难通过单次提示词（One-shot）直接交付高质量成果，通常需要反复提供反馈与多轮人工润色。为了衡量真实写作水平，测试引入了 Claude Opus 独立子代理充当客观评审，在没有任何历史对话上下文的前提下，将 GPT-6 Astra、Opus 5.5 与 Gemini 4 生成的初稿与创作者最终人工深度编辑的定稿进行盲测对比，考核维度包括：“个性化叙事风格（Personal）”、“洞察深刻度（Insightful）”以及“与最终成稿细节的契合度”。 ```mermaid flowchart TD accTitle: 访谈转 Newsletter 写作评估流程 accDescr: 原始访谈通过三款前沿模型生成初稿，并由独立 Opus 子代理参照人工定稿进行三维度盲测对比。 subgraph Input ["输入源数据"] A["原始访谈逐字稿"] end subgraph Generation ["模型初稿生成"] B1["Gemini 4 初稿"] B2["Opus 5.5 初稿"] B3["GPT-6 Astra 初稿"] end subgraph Evaluation ["无上下文 Opus 盲审"] C["评审维度： 1. 个人文风契合度 2. 洞察深度 3. 事实细节还原"] D["人工深度润色定稿"] end A --> B1 A --> B2 A --> B3 B1 --> C B2 --> C B3 --> C D -.->|"对齐基准"| C ``` 令人意外的是，Gemini 4 在该项测试中力压群雄夺得第一。评审结果明确指出，Gemini 4 是“唯一一份读起来真正像作者本人语气（Sounds like you）的文稿”，不仅准确摘录了访谈中最具力量的开场引言，还融入了鲜活的个人案例；虽然仍夹杂着极少量的行业术语，但整体语感与情绪饱满度最为出色。Opus 5.5 在洞察深度上得分最高，但在个性化风格上垫底；Astra 虽保证了事实准确，但在深度与人味上均表现平平。相较于前代 Gemini 3.8 容易假大空、啰嗦冗长的毛病，Google 显然对 Gemini 4 的长文本写作与语调进行了针对性的大幅调优。 --- ### [16:31](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=991s) Gemini 4 的隐藏杀招：深度医学研究与专业检索除了日常写作，Gemini 4 在严谨专业调研领域的突破更为惊人。由于母亲长期面临健康问题，创作者经常需要借助前沿 AI 进行复杂的医学文献调研与治疗方案梳理。在输入复杂的医学咨询场景后，Gemini 4 给出的诊断推导与方案极具针对性，逻辑严密且引证详实，挖掘出了其他前沿模型完全遗漏的关键医学信息。这项能力证明了模型的真实商业与生活价值。以往用户常常诟病 Gemini 产出的文本“过于详尽、篇幅过长（Verbose）”，但在面对关乎生命的医学调研或高门槛专业分析时，详尽与穷尽反而成为了决定性的竞争优势。结合其全新进化的写作语感，Gemini 4 在“专业研究（Research）”与“深度长文写作”两个维度上展现出了统治级的表现。 --- ### [17:45](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=1065s) 四大前沿模型横评大复盘与选型指南综合五大维度的严格实测，当前四大前沿模型并无绝对的单一霸主，而是呈现出鲜明的专长分化： | 模型代号 | 核心优势领域 | 典型表现与评测总结 | |:--- |:--- |:--- | | **Claude Opus 5.5** | 3D 建模、全代码动画生成、音视频剪辑 | 视觉与多媒体生成的绝对天花板，在姬路城与星际争霸小游戏测试中展现出顶级的审美与代码生成力，最为炫酷。 | | **GPT-6 Astra** | 现代化网页设计与界面落地页生成 | 排版干净利落，产品转化思维极佳，能够无缝结合原生图像生成完善页面视觉。 | | **Gemini 4** | 深度文本写作、复杂专业检索与医学调研 | 文风质感较前代大幅跃迁，兼具拟人化的语言风格与无出其右的信息穷尽能力，是知识工作者的核心生产力工具。 | | **Muse Wonder** | 个人 Agent 协同与工作流托管（待深测） | 在代码、视觉与网页单项测试中偏弱，其真正潜力预计体现在 Muse App 内的原生个人智能体联动中。 | 在当前的技术格局下，选择模型的关键取决于具体的应用场景：追求极致的多媒体与 3D 资产生成首选 Opus 5.5；搭建高转化落地页首选 GPT-6 Astra；而面对日常知识工作中权重最高的写作、提炼与专业调研，Gemini 4 则是目前综合实力最强的工作伙伴。 ## Notable quotes > [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) 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. — Peter Yang > [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) 简直就像一部正规的日本旅游宣传大片 — Peter Yang > [00:43](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=43s) 几代之前的落后模型产物 — Peter Yang > [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) 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. — Peter Yang > [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) 极力想要给老板留下深刻印象的勤奋员工 — Peter Yang > [03:46](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=226s) 宛如 Opus 3.1 时代的作品 — Peter Yang > [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) 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. — Peter Yang > [07:02](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=422s) 看起来简直就像是在对着植物小便 — Peter Yang > [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) 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. — Peter Yang > [11:20](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=680s) Learn AI by watching it built > [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) 在四款模型中它绝对是态度最积极、最想让你满意的那个 — Peter Yang > [12:26](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=746s) Make something of AI > [14:16](https://www.youtube.com/watch?v=zueuzhBkd9Y&t=856s) 唯一一份读起来真正像作者本人语气的文稿 — Claude Opus 独立子代理 ## Mentioned Gemini 4, Opus 5.5, GPT-6 Astra, Muse Wonder, Peter Yang, Google, Anthropic, OpenAI, Meta, Alex Wang, Himeji Castle (姬路城), Blender, Three.js, Star Fox, StarCraft, Behind the Craft, Sue, Gemini 3.8, Opus 3.1 ## Video description Copy my best AI workflows to save time and automate busywork: https://www.behindthecraft.com 🤝 Subscribe to my practical AI newsletter read by 150K+ professionals: https://creatoreconomy.so Gemini 4 just launched and the frontier model race is now a four-horse race. I tested Gemini 4, Opus 5.5, GPT-6 Astra, and Muse Wonder across five real use cases: 3D models, games, animated shorts, website design, and writing. What I cover: (00:00) Four frontier models build Himeji Castle in code (00:43) Test 1: Himeji Castle in Blender with music (03:46) Test 2: A Star Fox x StarCraft rail shooter (07:02) Test 3: An animated short that makes you feel something (11:20) Test 4: Which model is best at website design (14:16) Test 5: Testing each model's writing capabilities (16:31) The use case that really surprised me about Gemini 4 (17:45) Recapping the winners for each use case Prompts I used: Himeji Castle: "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." Star Fox x StarCraft: "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." Emotional short: "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." Website redesign: "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." Where to find me: X: https://x.com/petergyang LinkedIn: https://www.linkedin.com/in/petergyang/ Subscribe to this channel - more tutorials coming soon! [Source](https://www.youtube.com/watch?v=zueuzhBkd9Y)

Source quotes

[0:43] Peter Yang: 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. [0:43] Peter Yang: 简直就像一部正规的日本旅游宣传大片 [0:43] Peter Yang: 几代之前的落后模型产物 [3:46] Peter Yang: 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. [3:46] Peter Yang: 极力想要给老板留下深刻印象的勤奋员工 [3:46] Peter Yang: 宛如 Opus 3.1 时代的作品 [7:02] Peter Yang: 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. [7:02] Peter Yang: 看起来简直就像是在对着植物小便 [11:20] Peter Yang: 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. [11:20] Learn AI by watching it built [12:26] Peter Yang: 在四款模型中它绝对是态度最积极、最想让你满意的那个 [12:26] Make something of AI [14:16] Claude Opus 独立子代理: 唯一一份读起来真正像作者本人语气的文稿

Source quotes

[0:43] Peter Yang: 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. [3:46] Peter Yang: 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. [3:46] Peter Yang: Switch up the level to Terrans against the Zerg, and model all the Zerg properly in Blender. The Terran ship can be a Wraith. [7:02] Peter Yang: 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. [7:02] Peter Yang: peeing into a plant [11:20] Peter Yang: 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. [11:20] Learn AI by watching it built. [11:20] Peter Yang: B effort [12:26] Peter Yang: tries the hardest [12:26] Make something of AI [14:16] Claude Opus sub-agent: the only draft that sounds like you

Original transcript

[0:00] This is a video of Himiji Castle in [0:02] Japan during cherry blossom season made [0:05] by four different Frontier models [0:07] entirely using code. I'll tell you later [0:10] which one is Opus 5.5, GPD6 Astra and [0:14] the new Gemini 4 and Muse Wonder. You [0:17] know, the Frontier model race is no [0:19] longer just between OpenAI and [0:21] Anthropic. With today's release of [0:23] Gemini 4 and the launch of Muse Wonder [0:26] is now a four-horse race. I had early [0:30] access to all these models except Muse. [0:32] And in this video, I'm going to show you [0:34] how all four compare across five use [0:37] cases: 3D models, games, videos, website [0:42] design, and writing. All right, let's go [0:45] back to the castle. So, Himichi is one [0:47] of my favorite places in Japan. It's [0:50] white walls and layer roofs are really [0:52] hard for a model to get right, which [0:54] makes it a great test. So, here's the [0:56] prompt that I gave each model. Build [0:58] Himichi Castle in Blender and render a [1:01] 30 secondond cinematic at dawn in cherry [1:04] blossom season. Get the architecture [1:06] right and also add music all entirely [1:08] using code. All right, let's play the [1:11] best version first. [1:22] [music] [1:27] All right. So, this one is coming from [1:29] Opus 5.5, and I think it's pretty [1:32] incredible. I love the shots from the [1:35] cherry trees to the castle and also the [1:37] snow falling in the background. The [1:39] castle is also beautifully rendered [1:41] exactly like I remember it. It really [1:43] does look like a proper travel [1:45] commercial. So, Opus 5.5 is still [1:48] incredible. But now, let's take a look [1:49] at Gemini 4's version. [1:59] >> [music] [2:06] >> Yeah. So, I think this looks pretty [2:07] good, too. I think the castle also looks [2:09] like him. There's a lot more different [2:11] camera angles and shots like below the [2:13] bridge. And overall, it's beautifully [2:15] rendered. It's funny how like both [2:17] Gemini 4 and Opus have the flute music, [2:20] the traditional Japanese sounds. So, I [2:22] think Gemini 4 is a clear runner up. [2:25] Now, let's go ahead and play GPT Astra's [2:27] version. [2:35] [music] [2:43] Awesome. So, I really love the [2:45] reflections in the water and I can even [2:47] hear some sounds of birds, but I think [2:49] the castle doesn't look quite right. It [2:50] doesn't quite look like him. Looks a [2:52] little bit too flat. And also, the [2:54] camera angles aren't as interesting as [2:56] Opus and Gemini's version. But now, [3:00] let's take a look at Meta's Muse Wonder. [3:17] All right. All right. So, as you can [3:18] see, uh, Meta's Muse Wonder needs some [3:21] work. Now, to be fair, Alex Wang, who is [3:23] Meta's chief AI officer, said the model [3:25] is not trained on 3D models, but, uh, [3:28] this is by far obviously the worst. It [3:29] looks like, uh, a render from like a few [3:32] generations of models before. And yeah, [3:34] there's like a lot of room to improve [3:36] here. Awesome. So, that's the 3D model [3:38] test. Now, in my humble opinion, Opus is [3:41] the clear winner here with Gemini as a [3:43] runner up, Astra in number three, and [3:45] Muse as number four. Let's move on to [3:47] our next test, which is getting all four [3:49] models to build a game. So, each time a [3:52] new model comes out, I always like to [3:54] ask it to build Star Fox to see how it [3:56] is. This is the last one I did with GPT [3:59] 5.6 versus Faith Fable, but now let's [4:02] test this across all four models. And [4:04] this time, I did an interesting twist of [4:06] the prompt. I wanted to actually have it [4:08] build Star Fox style rail shooter with [4:12] Starcraft assets. Starcraft is probably [4:14] my number one favorite game. So, if you [4:16] have ever played before, you know it's [4:18] about Terrance, which is this human [4:20] race, defending themselves against the [4:21] Zerg, and there's also protos involved. [4:23] So, I wanted to have basically a Terran [4:26] ship shoot down a bunch of Zerg flyers. [4:30] Let's start with the best example, which [4:32] once again is Opus. So, let me go ahead [4:35] and play this game down. All right, I [4:37] put it on mute, but you know, I love the [4:39] camera angles. I love the music. I love [4:41] the sound effects. There's a little bit [4:42] too much lens flare for my taste. Uh, [4:45] too many things are shiny, but you know, [4:47] this is a pretty fun level that it [4:49] one-shotted. And here is a boss that [4:51] shoot at me. I can try to shoot it down [4:53] and so on so forth. And there's also a [4:55] chatter with my uh wingman on the bottom [4:58] left. So, yeah, this is pretty fun. Now, [5:02] let's go check out the next breast [5:03] version. Um, which I think is actually a [5:05] tie between Gemini and Astra. So, let me [5:08] show you Gemini's version first. Okay, [5:11] so this is Gemini's version. And right [5:13] off the bat, you can see that there's a [5:14] lot more features that Gemini built. So, [5:17] let's try playing it. All right. So, it [5:19] doesn't quite look as polished as Opus, [5:21] but let's try some of the new features, [5:22] right? So, I can press V, for example, [5:25] to go into firstperson view, which is [5:27] pretty unique. Um, I think down here I [5:30] can also, let's see, what else can I do? [5:33] I can press B to bring a battle cruiser [5:36] with a cannon. [laughter] It's pretty [5:38] fun. I can do barrel rolls, and I can do [5:41] more. I don't think this game looks as [5:43] polished as um Opus and it's a little [5:46] bit too crowded, but I'm pretty [5:48] impressed by all the features that [5:50] Gemini built in one shot. It seems like [5:52] the model likes to work very hard to try [5:54] to impress you. Awesome. So, now let's [5:57] go ahead and check out Astra's version. [6:00] Okay, so this is Astra's version and uh [6:03] it definitely has the moodiness and the [6:05] atmosphere that I'm looking for, but it [6:07] looks a little bit more basic. Let's go [6:09] ahead and play this. All right. So, as [6:11] you can see here, yeah, it's definitely [6:12] a lot more basic. There's only one enemy [6:15] type that I can see. The backgrounds [6:17] aren't as interesting. Uh, now that I [6:19] actually play both of these things, I [6:21] think even though Gemini's version is a [6:22] bit crowded, I think it's the clear [6:24] runner up to Opus. Astra is a little bit [6:26] too basic. Okay. Um, and as always, I [6:30] saved uh, you know, the funniest one for [6:34] last. So, let's go ahead and play the [6:35] one that Muse built. All right. Here we [6:38] are. This is the Musever version. Um, [6:42] yeah, this doesn't look that good, [6:44] right? This is looks um something that [6:46] maybe uh I don't know, Opus 3.1 built. [6:49] Again, Mus's defense is not really [6:52] trained for 3D models and games. Uh [6:55] maybe there's some other use cases like [6:56] personal agent that it does better than [6:58] the other models, but yeah, this this [7:00] isn't very impressive at all. Okay, so [7:03] for our third test, I wanted to see how [7:05] good these four models are at [7:07] storytelling by getting them to create [7:10] an animated emotional short that leaves [7:13] viewers with a strong emotion. And once [7:16] again, make every frame and add sound [7:18] all using code. So for this one, I think [7:21] the clear winner depends on your taste. [7:23] Let's go ahead and just play each one by [7:25] one. Starting with Opus. [7:37] All [7:52] >> [music] [7:57] >> right. So, from a graphics and animation [7:59] perspective, Opus, I think once again, [8:02] is the most impressive. The story line [8:04] seems to be about a wife waiting for her [8:06] husband to come back and leaving the [8:08] light on for him. I think it's okay. [8:10] Like, does it make me feel really [8:12] emotional? I'm not so sure. But yeah, [8:14] that is Opus. All right. Now, let's go [8:16] ahead and check out Gemini 4's version. [8:28] [music] [8:45] >> [music] [8:48] >> I think this is actually really [8:50] artistically well done about the doves [8:52] flying and the different scenes. I think [8:54] the story is about maybe someone who is [8:58] hard at work and hasn't really kept in [9:00] touch with their loved ones. That's my [9:02] impression at least. So, I think this is [9:04] very well done. It's very artistic. I [9:07] think I felt a little bit more emotion [9:09] with Opus to be honest with you. But [9:11] from a design and artistic point of [9:13] view, I think Gemini did a really great [9:14] job. All right. Now, let's check out [9:16] Astra's version. [9:20] [music] [9:39] >> [music] [9:46] [music] [9:47] >> Yeah, this is actually overall really [9:49] well done, too. I like the robot. It's [9:50] very cute. And I think the story is [9:53] about how a robot is watering the plant. [9:55] The plant grew higher and the sacrifices [9:57] that we make. For some reason, the robot [9:58] came back to life later on at the end of [10:00] the video. The only thing I don't really [10:02] like is this part. It seems like the [10:04] robot is like peeing into a plant or [10:06] something. It doesn't seem like it's [10:07] actually watering the plant. So, I don't [10:08] really know what's going on there. So, [10:11] in terms of emotional, I found this [10:12] pretty funny. I'm not so sure. It got me [10:14] very emotional. [10:16] All right. So, once again, I saved the [10:18] best or maybe the worst for last. Let's [10:19] check out the one from Muse. [10:22] [music] [10:28] >> [music] [10:44] [music] [10:50] >> Wow. You know what? I think Muse created [10:51] the most emotional video. At first, it's [10:54] just about a old man taking care of a [10:56] plant, but it turns out that he's taking [10:57] care of the plant so that he can grow a [10:59] beautiful flower and he can leave the [11:01] flower in the tombstone of I'm assuming [11:03] his spouse. The tombstone thing was a [11:05] surprise to me at the end and I think [11:07] Muse did the best job even though the [11:09] graphics aren't as good as Opus as some [11:11] of the other options. [11:13] So, yeah, I think for emotional short, [11:15] I'm actually have to give the award to [11:17] Muse here. Muse did a great job. [11:19] [snorts] Awesome. So the next thing I [11:21] want to show you guys is getting all [11:24] four models to redesign my website [11:27] without any context about what my [11:28] website is. So the prompt is redesign [11:31] the brand and homepage for behind the [11:33] craft, my newsletter, YouTube channel [11:35] and podcast. Create a logo, create a [11:37] color and typography and make a homepage [11:40] that doesn't look AI generated. Don't [11:42] look at my current site. Start from [11:43] scratch and let's take a look at what [11:45] each model did. So let's start with Opus [11:48] once again. So this is the website. I [11:51] think this is pretty cool. Pull the [11:53] corner to see what's behind it. It seems [11:55] like okay, learn AI by watching it [11:57] built. This is a pretty interesting [11:58] interaction pattern. The rest of it is [12:02] okay, but I don't know kind of boring. I [12:05] don't think it's that interesting. And [12:07] yeah, I think this is probably like a B [12:09] effort in my opinion. I wish there were [12:10] some images and other stuff, but again, [12:13] you know, claw doesn't actually have [12:14] access to image generation. Now, let's [12:17] take a look at Gemini's version. So, [12:20] this one, Gemini, unfortunately, didn't [12:22] seem to do a very good job. There's way [12:24] too much copy on here. It's like [12:26] extremely overwhelming how much stuff is [12:28] on here. It did capture my course and [12:31] everything else. But, I think it's just [12:33] like too much. It's too much copy and [12:34] too much stuff to read. And this is [12:36] actually similar to the game that Gemini [12:37] made. It build a ton of features and it [12:39] tried a lot, but as a result, the [12:42] website at least looks a little bit over [12:44] stuffed. So, I think Gemini can [12:46] definitely improve on this one. I do [12:48] like how he kind of built a design [12:50] system here with the different colors [12:52] and everything else and topography as [12:55] part of his effort. You know, Gemini, I [12:56] think out of all four models, tries the [12:59] hardest. It just keeps going and uh [13:01] tries to make you happy. But in this [13:02] case, I think it went a little bit too [13:04] overboard. All right, now let's look at [13:06] Astra's version. And yeah, I really like [13:10] Astra's version. I think it's very [13:12] simple. Make something of AI. Seems like [13:14] a good tagline. There's a clear call to [13:16] action to enter the email address. And [13:19] down here, I really like how Astra use [13:21] chatbt image generation to add images of [13:25] people I've interviewed, to add [13:26] illustrations here, and I like how [13:29] simple and clean it is. So, OpenAI isn't [13:33] exactly known for having the best design [13:35] sense for its models, but Astra really [13:38] did a good job. I think unless Muse [13:40] blows me away, I think Austra is the [13:41] winner for this one. Speaking of Muse, [13:44] this is what Muse made for me. [13:46] Unfortunately, I don't think Muse did a [13:48] good job here. I think it's extremely [13:50] boring. It's all text. It feels like [13:52] reading a newspaper. It doesn't feel [13:53] like a landing page. I think Muse didn't [13:55] do very well here. So, going back here, [13:58] I think for this one, I'm going to have [14:00] to give the war to Astra. I think [14:02] Astra's website design at least in this [14:03] particular example is the best followed [14:06] by Opus. And between Gemini and Muse, I [14:10] think they're in third place. Gemini [14:11] added a little bit too much and Muse [14:13] version looks extremely boring. Awesome. [14:16] So far, we've done three 3D model and [14:19] video tests and one website design test. [14:21] And now I wanted to move on to a much [14:23] more common use case that maybe is not [14:25] as flashy, which is how good each model [14:27] is at writing. So, I have a skill that [14:31] turns a raw interview transcript. This [14:33] is my interview with Sue into a [14:35] newsletter post that looks like this [14:37] with an introduction and a bunch of [14:40] takeaways down here with the top [14:42] insights from the interview. And despite [14:44] my best attempts, I have never been able [14:46] to actually update the skill to a point [14:48] where it can oneshot these takeaways and [14:51] make them good. I always have to provide [14:52] a feedback and go back and forth. What I [14:55] really wanted to do was I wanted to see [14:58] how close these Frontier models could [15:00] get to the final output here that I mali [15:04] helped edit. And for this one, I tested [15:07] Astra, Opus, and Gemini. I didn't get a [15:10] chance to test Muse. And here is the [15:13] verdict. So, the verdict is based on [15:15] asking Opus to review the copy that each [15:18] model generated and compare it with my [15:20] final copy on how personal it is, how [15:23] insightful it is, and how much details [15:26] matched the final copy. And surprisingly [15:28] here, Gemini actually came out first [15:31] versus Opus and Astra. And Gemini wins [15:34] because it is the only draft that sounds [15:36] like you. So, this is a Opus sub agent. [15:38] Doesn't have any context about any past [15:40] conversation. is purely just looking at [15:42] the output of these three different [15:44] bottles for my newsletter post versus my [15:46] final output. But basically, it's saying [15:48] Gemini is the only draft that sounds [15:50] like you is using the right opening [15:52] quote. It has personal example. It does [15:54] have some jargon still, but overall it [15:57] seems to be the best and opus was the [15:59] most insightful, but it scored the [16:01] lowest on personal and Astra is [16:04] accurate, but didn't score very high on [16:06] personal or insightful. [snorts] So, I'm [16:10] actually really surprised by Gemini 4's [16:12] improvement in writing. It's a lot [16:16] better than Gemini 3.8, the previous [16:19] version, because my understanding is [16:20] that the Google team actually spent a [16:22] lot of time to try to improve the [16:24] writing output. Gemini's output has [16:25] historically been like super verbose, [16:28] and right now I think it's a lot better. [16:31] And that actually extends to another use [16:32] case that I want to talk to you guys [16:34] about. So, my mom has been struggling [16:36] with some health issues and I've been [16:38] doing a lot of medical research with AI [16:40] to figure out, you know, how to get her [16:43] treatment going and so on and so forth. [16:45] And I was actually surprised by Gemini [16:48] 4's answers to my questions about [16:51] medical stuff. This is not the real [16:53] problem that she has, but this is just a [16:55] example. But, you know, if you look [16:57] through this result, I love how specific [17:00] and well researched it is. It found [17:03] things that the other frontier models [17:05] did not find. And you know, this is like [17:07] a real use case. This is something [17:09] that's really important to me versus [17:11] generating 3D games and 3D models. So, I [17:13] think really pleasantly surprised by how [17:16] good Gemini 4 did at doing medical [17:19] research, answering the questions that I [17:20] had and so on and so forth. I think [17:23] Gemini has always been the most [17:25] comprehensive and verbose model. [17:28] [snorts] [17:28] And in this case, it's actually playing [17:30] to its strengths because for something [17:32] like medical questions, I actually do [17:34] want something that's super well [17:35] researched. And that combined with its [17:37] improvement in writing makes this model [17:40] very good for writing and research. I [17:42] would actually give it first place in [17:44] these two areas. Okay. So, which model [17:47] wins? Well, I think it really comes down [17:49] to the use case. Opus is still best in [17:52] class on 3D models, animations, and you [17:56] know, editing videos and adding music [17:57] and all kind of stuff. Opus is probably [17:59] the most flashy right now. I think Astra [18:02] won on website design because it was the [18:04] most clean and is able to leverage [18:06] chatbt image generation, which to be [18:08] honest with you, I think Gemini can do [18:09] too with Nano Banana, but for some [18:11] reason it didn't. So hopefully Google [18:13] fixes that. And I think Gemini wins on [18:17] writing and research, which if you think [18:20] about it, are not as flashy, but are [18:22] actually really important tasks the [18:24] knowledge workers have to do. And then [18:26] when it comes to Muse, I think Muse [18:28] maybe is the best at helping a personal [18:30] agent with the Muse app. We'll have to [18:33] save that for a future video. Okay, so [18:36] I'm going to include all the prompts for [18:38] the use cases in the video description [18:40] if you want to try it out yourself. And [18:43] as always, if you like practical, [18:46] nononsense AI tutorials like this, [18:48] please like and subscribe. I'll see you [18:50] next time.