---
url: "https://youtu.be/zueuzhBkd9Y?is=g-n70q57kQ45NVwS"
title: I Tested Gemini 4 vs Opus 5.5 vs GPT-6 Astra vs Muse Wonder (5 Use Cases)
source_kind: youtube
author: Peter Yang
captured: "2026-10-21T19:33:39+00:00"
comment_tree: false
topics: [models, opus, gpt, games, 3d]
summary: A hands-on comparison of Gemini 4, Opus 5.5, GPT-6 Astra, and Muse Wonder across 3D modeling, games, videos, web design, and writing, showing Opus leading and Muse lagging.
status: ok
---

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

Channel: Peter Yang

## Transcript

[0s] This is a video of Himeji Castle in
[2s] Japan during cherry blossom season made
[5s] by four different frontier models
[7s] entirely using code.
[9s] I'll tell you later which one is Opus
[11s] 5.5, GPT-6 Astra, and the new Gemini 4
[15s] and Muse Wander.
[17s] You know, the frontier model race is no
[19s] longer just between OpenAI and
[21s] Anthropic. With today's release of
[23s] Gemini 4 and the launch of Muse Wander,
[26s] it's now a four-horse race.
[29s] I had early access to all these models
[31s] except Muse. And in this video, I'm
[34s] going to show you how all four compare
[36s] across five use cases,
[38s] 3D models, games, videos, website
[41s] design, and writing.
[44s] All right, let's go back to the castle.
[46s] So, Himeji is one of my favorite places
[48s] in Japan. It's white walls and layered
[51s] roofs are really hard for a model to get
[53s] right, which makes it a great test. So,
[56s] here's the prompt that I gave each
[57s] model. Build Himeji Castle in Blender
[60s] and render a 30-second cinematic at dawn
[63s] in cherry blossom season. Get the
[65s] architecture right and also add music
[67s] all entirely using code.
[70s] All right, let's play the best version
[71s] first.
[82s] >> [music]
[87s] >> All right, so this one is coming from
[89s] Opus 5.5. And I think it's pretty
[92s] incredible. I love the shots from the
[95s] cherry trees to the castle and also the
[97s] snow falling in the background. The
[99s] castle is also beautifully rendered
[101s] exactly like I remember it. It really
[103s] does look like a proper travel
[105s] commercial.
[106s] So, Opus 5.5 is still incredible. But
[109s] now, let's take a look at Gemini 4's
[110s] version.
[119s] >> [music]
[126s] >> Yeah, so I think this looks pretty good,
[128s] too. I think the castle also looks like
[129s] Himeji. There's a lot more different
[131s] camera angles and shots like below the
[133s] bridge.
[134s] And overall, it's beautifully rendered.
[136s] It's funny how like both Gemini 4 and
[138s] Opus have the flute music, the
[140s] traditional Japanese sounds. So, I think
[142s] Gemini 4 is a clear runner-up. Now,
[145s] let's go ahead and play GPT Astra's
[147s] version.
[155s] >> [music]
[163s] >> Awesome. So, I really love the
[165s] reflections in the water, and I can even
[167s] hear some sounds of birds. But, I think
[169s] the castle doesn't look quite right. It
[170s] doesn't quite look like Himeji. Looks a
[172s] little bit too flat. And also, the
[173s] camera angles aren't as interesting as
[176s] Opus and Gemini's version.
[179s] But now, let's take a look at Meta's
[181s] Muse Wonder.
[197s] All right. So, as you can see, uh Meta's
[199s] Muse Wonder needs some work. Now, to be
[201s] fair, Alex Wang, who is Meta's chief AI
[204s] officer, said the model is not trained
[205s] on 3D models.
[207s] But, uh this is by far obviously the
[209s] worst. It looks like
[210s] a a render from like a few generations
[212s] of models before.
[214s] And yeah, there's like a lot of room to
[215s] improve here.
[217s] Awesome. So, that's the 3D model test.
[219s] Now, in my humble opinion, Opus is the
[221s] clear winner here with Gemini as the
[223s] runner-up, Astra in number three, and
[225s] Muse as number four. Let's move on to
[227s] our next test, which is getting all four
[229s] models to build a game.
[231s] So, each time a new model comes out, I
[233s] always like to ask it to build Starfox
[236s] to see how it is. This is the last one I
[238s] did, which is GPT-5.6 versus Fai Fable,
[241s] but now let's test this across all four
[243s] models. And this time I did an
[245s] interesting twist with the prompt. I
[247s] wanted to actually have it build
[249s] Starfox-style rail shooter
[251s] with Starcraft assets. Starcraft is
[253s] probably my number one favorite game.
[255s] So, if you have ever played before, you
[258s] know it's about Terrans, which is this
[259s] human race defending themselves against
[261s] the Zerg, and there's also Protoss
[262s] involved. So, I want to have basically a
[265s] Terran ship shoot down a bunch of Zerg
[268s] flyers.
[270s] Let's start with the best example, which
[272s] once again is Opus.
[274s] So, let me go ahead and play this game
[275s] now.
[277s] All right, I put it on mute, but you
[278s] know, I love the camera angles, I love
[280s] the music, I love the sound effects.
[281s] There's a little bit too much lens flare
[283s] for my taste. Uh
[285s] to me things are shiny, but you know,
[287s] this is a pretty fun level that it
[289s] one-shotted. And here is a boss that's
[291s] shooting at me. I can try to shoot it
[292s] down, and so on and so forth. And
[294s] there's also a chatter with my uh
[297s] wingman on the bottom left.
[299s] So, yeah, this is pretty fun.
[301s] Now, let's go check out the next best
[303s] version,
[304s] um which I think is actually a tie
[306s] between Gemini and Astra. So, let me
[308s] show you Gemini's version first.
[311s] Okay. So, this is Gemini's version. And
[313s] right off the bat you can see that
[314s] there's a lot more features that Gemini
[316s] built. So, let's try playing it.
[318s] All right, so it doesn't quite look as
[320s] polished as Opus, but let's try some of
[322s] the new features, right? So,
[323s] I can press V, for example, to go into
[325s] first-person view, which is pretty
[327s] unique.
[328s] Um I think down here I can also, let's
[331s] see, what else can I do? I can press B
[334s] to bring a battlecruiser with a cannon,
[337s] >> [laughter]
[338s] >> which is pretty fun. I can do barrel
[340s] rolls,
[341s] and I can do more. I don't think this
[342s] game looks as polished as um
[345s] Opus. It's a little bit too crowded, but
[348s] I'm pretty impressed by all the features
[350s] that Gemini built in one shot. It seems
[352s] like the model likes to work very hard
[353s] to try to impress you.
[355s] Awesome. So, now let's go ahead and
[357s] check out Astra's version.
[360s] Okay, so this is Astra's version and uh
[363s] it definitely has the moodiness and the
[365s] atmosphere that I'm looking for, but it
[367s] looks a little bit more basic. Let's go
[369s] ahead and play this.
[370s] All right, so as you can see here, yeah,
[372s] it's definitely a lot more basic.
[373s] There's only one enemy type that I can
[375s] see. The backgrounds aren't as
[377s] interesting.
[378s] Uh now that I actually played both of
[380s] these things, I think even though
[381s] Gemini's version is a bit crowded, I
[383s] think it's the clear runner-up to uh
[385s] Opus. Astra's a little bit too basic.
[388s] Okay.
[388s] Um, and as always, I saved uh
[392s] you know,
[392s] the funniest one for last. So, let's go
[395s] ahead and play the one that Muse built.
[398s] All right, here we are. This is the Muse
[399s] version.
[400s] Um
[402s] Yeah, this doesn't look that good,
[404s] right? This just looks um something that
[405s] maybe uh
[407s] I don't know, Opus Super 1 built.
[409s] Again, in Muse's defense, it's not
[412s] really trained for 3D models and games.
[414s] Uh maybe there's some other use cases
[416s] like personal agent that it does better
[417s] than the other models.
[419s] But yeah, this this isn't very
[420s] impressive at all. Okay, so for our
[423s] third test, I wanted to see how good
[426s] these four models are at storytelling by
[428s] getting them to create an animated
[431s] emotional short that leaves viewers with
[434s] a strong emotion. And once again, make
[436s] every frame and add sound all using
[439s] code.
[440s] So, for this one, I think the clear
[442s] winner depends on your taste. Let's go
[444s] ahead and just play each one by one.
[446s] Starting with Opus.
[472s] >> [music]
[476s] >> All right. So, from a graphics and
[479s] animation perspective, Opus, I think
[481s] once again,
[482s] is the most impressive. The storyline
[484s] seems to be about a wife waiting for her
[486s] husband to come back and leaving the
[488s] light on for him.
[490s] I think it's okay. Like, does it make me
[491s] feel really emotional? I'm not so sure.
[493s] But, yeah. That is Opus.
[495s] All right. Now, let's go ahead and check
[496s] out Gemini 4's version.
[508s] >> [music]
[525s] [music]
[528s] >> I think this is actually really
[529s] artistically well done about the doves
[532s] flying and the different scenes. I think
[534s] the story is about maybe someone who is
[537s] hard at work and hasn't really kept in
[539s] touch with their loved ones.
[541s] That's my impression at least.
[543s] So, I think this is very well done. It's
[545s] very artistic.
[547s] I think I felt a little bit more
[548s] emotional with Opus, to be honest with
[549s] you.
[550s] But, from a designer artistic point of
[552s] view, I think Gemini did a really great
[554s] job.
[555s] All right. Now, let's check out Astra's
[557s] version.
[560s] >> [music]
[579s] [music]
[586s] [music]
[587s] >> Yeah, this is actually overall really
[588s] well done, too. I like the robot. It's
[590s] very cute. And I think the story is
[592s] about how a robot is watering the plant.
[595s] The plant grew higher, and the
[596s] sacrifices that we make. For some
[598s] reason, the robot came back to life
[599s] later on at the end of the video.
[602s] The only thing I don't really like is
[603s] this part. It seems like the robot is
[604s] like peeing into the plant or something.
[606s] It doesn't seem like it's actually
[607s] watering the plant. So, I don't really
[608s] know what's going on there.
[610s] So, in terms of emotional, I thought
[612s] it's pretty funny. I'm not so sure it
[614s] got me very emotional.
[616s] All right. So, once again, I saved the
[618s] best or maybe the worst for last. Let's
[619s] check out the one from Mews.
[622s] >> [music]
[628s] [music]
[644s] [music]
[650s] >> Well, you know what? I think Mews
[651s] created the most emotional video.
[653s] At first, it's just about an old man
[655s] taking care of a plant. But, it turns
[657s] out that he's taking care of the plant
[658s] so that he can grow a beautiful flower,
[660s] and he can leave the flower in the
[662s] tombstone of I'm assuming his spouse.
[664s] The tombstone thing was a surprise to me
[666s] at the end. And I think Mews did the
[668s] best job, even though the graphics
[669s] aren't as good as Opus and some of the
[671s] other options.
[673s] So, yeah, I think for emotional short,
[675s] I'm actually going to give the award to
[677s] Mews here. Mews did a great job.
[679s] >> [snorts]
[679s] >> Awesome. So, the next thing I want to
[681s] show you guys is getting all four models
[684s] to redesign my website
[687s] without any context about what my
[688s] website is. So, the prompt is, "Redesign
[691s] the brand and homepage for behind the
[693s] craft, my newsletter, YouTube channel,
[695s] and podcast. Create a logo, create a
[697s] color and typography, and make a home
[699s] page that doesn't look AI generated.
[701s] Don't look at my current site, start
[703s] from scratch.
[704s] And let's take a look at what each model
[706s] did.
[707s] So, let's start with Opus once again.
[710s] So, this is the website. I think this is
[711s] pretty cool. Pull the corner to see
[713s] what's behind it.
[715s] Seems like, okay, learn AI by watching
[717s] it build. This is pretty interesting
[718s] interaction pattern.
[720s] The rest of it is okay,
[723s] but I don't know, kind of boring. I
[725s] don't think it's that interesting.
[726s] And yeah, I think this is probably like
[728s] a B effort, in my opinion. I wish there
[730s] were some images and other stuff, but
[732s] again, you know, Claude doesn't actually
[733s] have access to image generation.
[736s] Now, let's take a look at Gemini's
[738s] version.
[739s] So, this one, Gemini, unfortunately,
[742s] didn't seem to do a very good job.
[744s] There's way too much copy on here. It's
[745s] like extremely overwhelming how much
[747s] stuff is on here.
[749s] It did capture my course and everything
[751s] else,
[752s] but I think it's just like too much.
[753s] There's too much copy and too much stuff
[754s] to read. And this is actually similar to
[756s] the game that Gemini made. It built a
[758s] ton of features, and it tried a lot, but
[761s] as a result, the website, at least,
[763s] looks a little bit overstuffed.
[765s] So, I think Gemini can definitely
[766s] improve on this one.
[768s] I do like how it kind of built a design
[769s] system here
[771s] with the different colors and everything
[773s] else and typography
[774s] as part of his effort. You know, Gemini,
[776s] I think, out of all four models,
[778s] tries the hardest. It is keeps going and
[780s] uh tries to make you happy. But in this
[782s] case, I think it went a little bit too
[784s] overboard.
[785s] All right. Now, let's look at Astra's
[787s] version.
[788s] And yeah, I really like Astra's version.
[791s] I think it's very simple. Make something
[793s] of AI. Seems like a good tagline.
[795s] There's a clear call to action to enter
[797s] the email address.
[798s] And down here, I really like how Astra
[801s] used ChatGPT image generation to add
[804s] images of people I've interviewed, to
[806s] add illustrations here.
[808s] And I like how simple and clean it is.
[811s] So,
[812s] OpenAI isn't exactly known for having
[814s] the best design sense for its models,
[817s] but Astra really did a good job, I
[818s] think. Unless Muse blows me away, I
[820s] think Astra's the winner for this one.
[823s] Speaking of Muse, this is what Muse made
[825s] for me.
[826s] Unfortunately, I don't think Muse did a
[828s] good job here. I think it's extremely
[830s] boring. It's all text. It feels like
[831s] reading a newspaper. It doesn't feel
[833s] like a landing page.
[834s] I think Muse didn't do very well here.
[837s] So, going back here, I think for this
[839s] one, I'm going to have to give the award
[840s] to Astra. I think Astra's website
[842s] design, at least in this particular
[844s] example, is the best, followed by Opus.
[847s] And between Gemini and Muse, I think
[850s] they're in third place. Gemini added a
[851s] little bit too much, and Muse's version
[854s] looks extremely boring.
[856s] Awesome. So, so far we've done three 3D
[858s] model and video tests and one website
[860s] design test. And now I want to move on
[862s] to a much more common use case that
[864s] maybe it's not as flashy, which is how
[866s] good each model is at writing. So, I
[870s] have a skill that turns a raw interview
[872s] transcript. This is my interview with
[873s] Sue.
[874s] Into a newsletter post that looks like
[876s] this, with introduction and a bunch of
[880s] takeaways down here with the top
[881s] insights from the interview. And despite
[884s] my best attempts, I have never been able
[886s] to actually update this skill to a point
[888s] where it can one-shot these takeaways
[891s] and make them good. I always have to
[892s] provide a feedback and go back and
[893s] forth.
[894s] What I really want to do was I wanted to
[896s] see how close these frontier models
[899s] could get
[901s] to the final output here that I manually
[904s] helped edit. And for this one, I tested
[907s] Astra, Opus, and Gemini. I didn't get a
[909s] chance to test Muse.
[912s] And here is the verdict. So, the verdict
[914s] is based on asking Opus to review the
[917s] copy that each model generated and
[919s] compare it with my final copy
[922s] on how personal it is, how insightful it
[924s] is, and how much details matched the
[926s] final copy. And surprisingly here,
[928s] Gemini actually came out first versus
[931s] Opus and Astra.
[933s] And Gemini wins because it is the only
[935s] draft that sounds like you. So, this is
[937s] the Opus sub agent doesn't have any
[939s] context about any past conversation.
[941s] It's purely just looking at the output
[943s] of these three different models for my
[945s] newsletter post versus my final output.
[947s] But basically it's saying Gemini is the
[949s] only draft that sounds like you, is
[951s] using the right opening quote, it has
[953s] personal example. It does have some
[955s] jargon still, but overall it seems to be
[957s] the best. And Opus was the most
[959s] insightful, but it scored lowest on
[962s] personal. And Astra is accurate, but
[965s] didn't score very high on personal or
[967s] insightful.
[968s] >> [snorts]
[968s] >> So,
[970s] I'm actually really surprised by Gemini
[971s] 4's improvement in writing. It's a lot
[976s] better than Gemini 3.8, the previous
[979s] version, because my understanding is
[980s] that the Google team actually spent a
[981s] lot of time to try to improve the
[983s] writing output. Gemini's output has
[985s] historically been like super verbose,
[988s] and right now I think it's a lot better.
[991s] And that actually extends to another use
[992s] case that I want to talk to you guys
[994s] about. So, my mom has been struggling
[996s] with some health issues, and I've been
[998s] doing a lot of medical research with AI
[1000s] to figure out, you know, how to get her
[1003s] treatment going and so on and so forth.
[1005s] And I was actually surprised by Gemini
[1007s] 4's answers to my questions about
[1011s] medical stuff.
[1012s] This is not the real problem that she
[1014s] has, but this is just a example.
[1016s] But, you know, if you look through this
[1017s] result,
[1019s] I love how specific and well researched
[1021s] it is. It found things that the other
[1024s] frontier models did not find.
[1027s] And you know, this is like a real use
[1028s] case. This is something that's really
[1029s] important to me versus generating 3D
[1031s] games and 3D models. So, I think really
[1034s] pleasantly surprised by how good Gemini
[1037s] 4 did at doing medical research,
[1039s] answering the questions that I had, and
[1041s] so on and so forth.
[1043s] I think Gemini has always been the most
[1045s] comprehensive and verbose model.
[1048s] >> [snorts]
[1048s] >> And in this case, it's actually playing
[1049s] to its strengths.
[1051s] Because for something like medical
[1053s] questions, I actually do want something
[1054s] that's super well researched. And that
[1056s] combined with its improvement in writing
[1059s] makes this model very good for writing
[1062s] and research. I will actually give it
[1063s] first place in these two areas.
[1066s] Okay, so which model wins? Well, I think
[1068s] it really comes down to the use case.
[1070s] Opus is still best in class on 3D
[1073s] models, animations, and, you know,
[1076s] editing videos and adding music and all
[1077s] kinds of stuff.
[1079s] Opus is probably the most flashy right
[1080s] now.
[1081s] I think Astra won on website design
[1084s] because it was the most clean and it's
[1085s] able to leverage ChatGPT image
[1087s] generation, which to be honest with you,
[1088s] I think Gemini can do, too, with Nano
[1090s] Banana, but for some reason it didn't.
[1092s] So, hopefully Google fixes that.
[1094s] And I think Gemini wins on writing and
[1098s] research, which, if you think about it,
[1100s] are not as flashy, but are actually
[1102s] really important tasks that knowledge
[1104s] workers have to do. And then when it
[1106s] comes to Muse, I think Muse maybe is the
[1109s] best at helping personal agent with the
[1111s] Muse app. We'll have to save that for a
[1114s] future video.
[1116s] Okay, so I'm going to include all of the
[1118s] prompts for the use cases in the video
[1119s] description if you want to try it out
[1121s] yourself.
[1122s] And as always,
[1125s] if you like practical, no-nonsense AI
[1127s] tutorials like this, please like and
[1129s] subscribe. I'll see you next time.