If you pick a model by leaderboard rank, this test will frustrate you: Gemini 4 vs Opus 5.5 came down to task type. Opus produced the most polished 3D and game output, while Gemini produced the draft closest to the creator's own newsletter voice and the more thorough medical research answers.
Gemini 4 vs Opus 5.5: what the five-use-case test actually measured
Gemini 4 vs Opus 5.5 was decided per use case, not overall, in a test Peter Yang published on 13 October 2026 comparing four models on 3D scenes, games, animated shorts, website design and writing. No model won more than two categories, and the writing verdict came from an Opus model grading drafts rather than from blind human judges.
Yang, a creator who publishes AI tutorials on his YouTube channel, said he had early access to Gemini 4, Opus 5.5 and GPT-6 Astra but not to Muse Wonder, which he tested as a public release. The prompts asked each model to produce work entirely in code: Blender renders for a castle scene, a three.js rail shooter, a 30-second animated short, a homepage redesign for his newsletter and podcast, and a newsletter post from a raw interview transcript.
Two limits matter before reading any ranking. First, every verdict is one reviewer's judgment of four outputs per category, with no repeat runs, so a single lucky or unlucky generation could move a placement. Second, Muse Wonder was entered in categories its own launch framing did not target, which Yang acknowledges when he says the model was not trained on 3D models or games.
How the four models compare across the five tests
The comparison below reproduces Yang's stated placements only. Each cell records one test run per model per category, judged by a single reviewer, so treat it as a starting hypothesis for your own evaluation rather than a benchmark score.
| Category | Opus 5.5 | Gemini 4 | GPT-6 Astra | Muse Wonder |
|---|---|---|---|---|
| Himeji Castle render | 1st | 2nd | 3rd | 4th |
| Star Fox-style game | 1st | 2nd | 3rd (tie with Gemini in Yang's words) | 4th |
| Animated short | Strongest visuals | Most artistic | Cute but confusing | Most emotional |
| Website redesign | 2nd | 3rd (tie) | 1st | 3rd (tie) |
| Writing / research | Most insightful, least personal | 1st | 3rd | Not tested |
Yang's own summary splits the result by workflow: Opus for 3D models, animation and video work; Astra for clean web design backed by ChatGPT's image generation; Gemini for writing and research. He names no overall champion, and the transcript gives no reason to expect one model to sweep a task it was not built for.
Why Opus 5.5 won the 3D and game tests
Opus 5.5 won the 3D and game categories because its outputs looked most like finished commercial work: a dawn Himeji Castle render with cherry blossoms and falling snow, and a Star Fox-style rail shooter with a boss fight, wingman chatter and sound effects generated in a single prompt.
The 3D prompt asked for Himeji Castle in Blender with a 30-second cinematic at dawn in cherry blossom season, correct architecture, and music composed in code. Yang called the Opus result a proper travel commercial and praised the camera moves from cherry trees to the castle. Gemini's version also read as Himeji to him and used more camera angles, including a shot from under the bridge, which made it the clear runner-up.
In the game test, the prompt mixed two franchises: a Star Fox-style rail shooter with StarCraft assets, a Terran Wraith ship against Zerg flyers modeled in Blender. Opus produced the most polished level in one shot, with music, sound effects and a boss encounter. Yang's only complaint was too much lens flare. Gemini shipped more features, including a first-person view, a summonable battlecruiser and barrel rolls, but he described the result as crowded, and Astra's version as too basic with one enemy type.
Why Muse Wonder trailed on 3D and games
Muse Wonder finished last in both 3D and games, and the fairest reading is scope: Yang reports that Meta's chief AI officer said the model was not trained on 3D models, so the castle and shooter tests measured tasks outside its stated target.
The outputs were visibly weaker. Yang compared the Muse castle render to work from several model generations earlier, and the Muse shooter to something a much older model might build. He suggests Muse may be aimed at a personal-agent role inside the Muse app and says that comparison is left for a future video, which means this test says nothing about the category Muse was built for.
The animated short: best visuals versus strongest story
The animated short produced the one category where Yang gave the award against the visual ranking: Muse Wonder's old-man-and-flower story landed the emotional beat, while Opus 5.5 produced the most impressive graphics and Gemini 4 the most artistic imagery.
The prompt asked for a 30-second short that leaves viewers with a strong emotion, with every frame and sound made in code and no stock assets. Yang read Opus's film as a wife leaving a light on for her husband. Gemini's used doves and office scenes and read to him as someone out of touch with loved ones. Astra's featured a robot watering a plant, which he found cute but confusing in one shot. Muse's story turned out to be an old man growing a flower to place on a tombstone, a reveal Yang called the most emotional moment in the set despite weaker graphics.
That split is the useful signal. Visual fidelity and narrative payoff came from different models on the same prompt, so a short-video workflow may need one model for frames and another pass, or a different prompt, for story.
Why Astra won website design and Gemini overbuilt
GPT-6 Astra won website design because its redesign was the cleanest: a short tagline, a clear email signup, and images of real interviewees added through ChatGPT's image generation, which no other entry matched in polish.
All four models were asked to redesign the brand and homepage for Yang's newsletter, YouTube channel and podcast, create a logo, pick colors and typography, and avoid looking AI-generated, without looking at his current site. He rated Opus's result a B effort for a clever reveal interaction and a dull remainder, and said Claude no image generation to fill it out. Gemini built the most complete design system but buried the page in copy, echoing the crowded game. Muse produced plain text that Yang said read like a newspaper rather than a landing page.
The pattern across the two prior categories repeats here: Gemini attempts the most, Astra ships the most presentable, and Opus produces the most technically ambitious but least complete page.
Writing and research: Gemini 4's strongest result
Gemini 4 won the writing test because it produced the only draft that, in an Opus-graded comparison, sounded like the creator: it used the right opening quote and a personal example, though some jargon remained.
The task fed a raw interview transcript into a skill that turns it into a newsletter post with an introduction and takeaways, compared against Yang's manually edited final version. Because he had never gotten the takeaways to work in one shot, he asked another model to judge how personal, insightful and accurate each draft was. That judge scored Gemini first, Opus most insightful but weakest on personal voice, and Astra accurate but low on both.
Yang also described Gemini 4 as better at medical research questions than earlier Gemini versions, finding sources the other models missed, and said the writing improvement followed deliberate work by the Google team on verbosity. The research comparison is his impression across real questions, not a scored test, so treat it as a signal to try Gemini on literature-style queries rather than a measured lead.
FAQ
- Which model won the overall Gemini 4 vs Opus 5.5 test? There was no overall winner. Peter Yang split the result by task in his 13 October 2026 video: Opus for 3D, animation and games, Astra for website design, and Gemini for writing and research.
- Did Muse Wonder really lose every category? Muse finished last in 3D and games but won the animated short on emotional impact. Yang notes the model was not trained on 3D models, so two of those losses fall outside its intended use.
- Is the writing result a fair benchmark? It is a single comparison judged by an Opus model against the creator's edited copy, with one draft per model. It shows Gemini's draft matched his voice best in that run, not that Gemini writes better in general.
- What should I do with these rankings? Rerun the prompts Yang published in the video description on your own tasks and judge the outputs yourself. One generation per model per category is too thin to settle a purchase decision.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits