Intel·ligència artificial Models de llenguatge Anthropic Claude Code programació amb IA Claude Opus 4.8 Bijan Bowen

Claude Opus 4.8 sota prova: resultats brillants, però no supera clarament Opus 4.7

Bijan Bowen sotmet Claude Opus 4.8 a jocs, webs, animacions i simulacions 3D. Dos resultats són extraordinaris, però la resta mostren molta irregularitat.

Claude Opus 4.8 va arribar el 28 de maig del 2026 com a nou model de màxima gamma d’Anthropic. Bijan Bowen no es limita als gràfics de benchmarks: durant unes cinc hores el posa a construir un sistema operatiu de navegador, jocs en 3D, una animació SVG, simuladors i una web a partir d’imatges de referència.

La conclusió és incòmoda per a un llançament insígnia. Opus 4.8 produeix dos dels resultats més impressionants que el creador ha vist, però en diverses proves habituals queda per sota del seu Opus 4.7. No hi ha evidència suficient per proclamar-lo «el millor model»; sí que hi ha una capacitat enorme quan disposa de temps, eines i una tasca exigent.

Novetats tècniques, esforç i preu

Anthropic presenta Opus 4.8 com una millora tangible, però moderada, respecte del predecessor. El model incorpora pensament adaptatiu i un control d’esforç que permet decidir quants recursos dedicar a una tasca. Bijan alterna nivells alt, extra i màxim perquè el comportament canvia segons aquesta configuració.

El preu estàndard de l’API es manté en 5 dòlars per milió de tokens d’entrada i 25 per milió de sortida. El mode ràpid, que Anthropic anuncia amb una velocitat superior a canvi d’un recàrrec, puja a 10 i 50 dòlars respectivament. Això importa en fluxos llargs i amb subagents: la capacitat de continuar treballant durant molta estona també pot consumir pressupost amb rapidesa.

El vídeo comenta igualment els fluxos dinàmics de Claude Code i l’anunci d’una futura classe de models per sobre d’Opus. Són elements de la notícia de llançament; la prova, però, es concentra en el model disponible aquell dia i no en promeses futures.

Un sistema operatiu vistós, però incomplet

La primera prova demana un «browser OS» amb aplicacions, fons personalitzable i dos jocs 3D, un dels quals ha de recordar GTA. Opus 4.8 tria sense que se li demani una estètica synthwave molt coherent. Inclou cercador, Paint, calculadora, bloc de notes, canvi de tema, fons animat i un joc espacial.

El clon de GTA falla al primer intent perquè el cotxe apareix atrapat dins d’un edifici. Després d’una correcció, es pot conduir, baixar del vehicle, col·lidir amb policies i prendre un altre cotxe. Els models 3D són bons, però les finestres no es poden redimensionar ni posar a pantalla completa i tampoc hi ha menú contextual.

És un resultat atractiu, però la comparació directa afavoreix Opus 4.7: en aquella versió hi havia trànsit autònom i els personatges reaccionaven millor als impactes. L’exemple introdueix el patró de tota la sessió: 4.8 té una presentació potent, però no sempre conserva el detall funcional.

Animació SVG i joc de metro

La generació d’una peça publicitària animada en SVG surt bé. El model combina un guió de veu, actius gràfics i colors de marca en una seqüència fluida. Bijan veu aquí un ús pràctic: crear recursos de vídeo lleugers per a anuncis o xarxes sense recórrer a un editor tradicional per a cada moviment.

El següent repte converteix una estació de metro en un joc de trets. La il·luminació fluorescent, els reblons metàl·lics, els enemics i l’animació de l’arma tenen molt detall. El joc funciona amb fluïdesa i fins i tot inclou un interruptor per desactivar el parpelleig, una decisió d’accessibilitat que el prompt no exigia.

Tot i així, les bales no deixen marques a l’entorn. Opus 4.7 sí que havia generat aquests impactes en una prova comparable, a més d’un tren i una ambientació més completa. La nova versió és bona en termes absoluts, però no representa una millora lineal.

Skate en C++, apartament i simulacions irregulars

El joc de monopatí en C++ necessita una espera llarga sense indicar clarament si el model continua treballant. El resultat final és net, amb passeig marítim californià, aigua i trucs, però la versió anterior havia produït una escena visualment més rica després d’una iteració addicional.

La recreació de l’apartament de Jerry Seinfeld tampoc convenç al principi. En revisar la comparació, Bijan admet un matís important: havia donat una fotografia de referència a Opus 4.7 i no a 4.8. Això fa que el cara a cara no sigui controlat i evita extreure una conclusió contundent d’aquell resultat.

La simulació de vol és una de les pitjors proves. La primera versió funciona a molt pocs fotogrames per segon; després de demanar una simplificació, milloren els models d’avió però apareixen moviment cap enrere i una pantalla negra que no acaba. També hi ha bloquejos i errors de connexió quan el model pensa durant molta estona.

El simulador d’impressora 3D és més prometedor: la gantry es mou, la bobina gira i les capes es dibuixen. El problema és que un quadrat acaba semblant una forma circular i un triangle recorda més una gota. La mecànica és detallada, però l’objecte produït no respecta la geometria demanada.

Omni Racer: la redempció més difícil

Davant dels resultats irregulars, Bijan prepara un repte molt més exigent en mode ràpid i amb esforç màxim. Dona al model una fotografia d’una màquina recreativa i un full de deu cotxes generats amb IA. Li demana construir una web 3D, modelar la cabina i crear un joc de curses jugable a la seva pantalla amb aquells sprites.

El full no té un format net, de manera que el model ha d’aïllar els cotxes, convertir-los en recursos i trobar una manera de crear profunditat. També ajusta proporcions, cargols, volant, palanca i il·luminació després d’observar captures del seu propi resultat.

Al cap d’aproximadament una hora i mitja, la recreativa funciona. El joc respon a les tecles WASD, el volant i la palanca s’animen i els cotxes procedents de la imatge queden integrats en una carretera 3D. La cabina no és una còpia exacta i hi ha detalls visuals millorables, però la combinació d’anàlisi d’imatge, modelatge, codi i iteració és extraordinària.

Una bateria virtual també sorprenent

L’última prova genera una bateria completa amb peces animades, so, seqüenciador i reproducció automàtica de quatre ritmes. Els plats tenen curvatura, els cops mostren una resposta visual clara i es pot canviar el tempo. El model arriba al límit de longitud durant el procés, però el producte resultant funciona.

Omni Racer i la bateria demostren que Opus 4.8 pot coordinar projectes llargs amb moltes parts. Alhora, requereixen més temps i consum que una resposta convencional. El millor rendiment apareix quan el model pot explorar, executar, mirar el resultat i corregir-lo, no necessàriament al primer intent.

Veredicte: molt capaç, però massa inconsistent

Bijan acaba dividit. Les dues proves finals són fantàstiques, i Omni Racer és possiblement un dels projectes més impressionants del seu canal. En canvi, el navegador, el metro, el joc de skate, la simulació de vol i altres resultats no superen de manera clara el model anterior; alguns són directament pitjors.

La configuració també impedeix una sentència universal. L’esforç, el pensament adaptatiu, el mode ràpid, les referències visuals i el nombre d’iteracions alteren molt el resultat. Una prova manual no equival a un benchmark científic, però sí que exposa una realitat útil: una mitjana millor no garanteix que cada projecte millori.

El resum més precís és que Claude Opus 4.8 eleva el sostre, no necessàriament el terra. Pot completar treballs multimodals i interactius de gran complexitat, però el seu comportament inicial és irregular i car de validar. El vídeo no el corona com el millor model; recomana jutjar-lo amb les tasques reals de cada usuari.

Contrast i context

Fonts consultades

4 fonts
  1. 01
  2. 02
  3. 03
  4. 04

Font de treball

Transcripció amb marques de temps

10 fragments
Consulta la transcripció
  1. 0:00 , obre el vídeo en una pestanya nova

    Oh my, what the heck? What sort of squares does this sick even from here like from this view as well? Camera angles that we could see. Oh and Thropic has released cloud opus 4.8 which is quite exciting because this is a new frontier state of the art model which always is exciting I guess could be said. Now before we get into this please do feel free to subscribe as I do want that 6 figure YouTube plaque but we're gonna take a quick look at this truthfully relatively short introductory post which I am okay with because then we can just jump straight into testing the model though I do have to say there are some pretty interesting things contained in this short blog post that I do want to touch upon because some of them talk about future potential models that would actually sit a class above the Opus family, which is still the highest and publicly available anthropic model. So with that let's take a look at this benchmark JPEG, where we see right here that it is beating out its predecessor in every specific. category right here the only model in this chart that it is not beating in one specific area is GPT55 and Gemini 3.1 pro is pretty old that this point and will be replaced this month with Gemini 3.5 pro so it'll be interesting to see how these scores change over time now I don't pay too much mind to these benchmark jpegs really and there has been a new benchmark released where apparently it was a private one that tested these encoding specific things and the scores changed drastically a lot of model were performing significantly worse than they were on some other public like well-known benchmarks. I think five-five was ahead. That's something that there's probably a lot of talk about if you're interested in such a thing. No something else that's kind of I won't say concerning, but it was like, oh, okay. So they talk about the model is reaching new heights on our measures of pro social traits like supporting user autonomy and then this one acting in the user's best interest. That will go kind of concerns me and that may be a result of seeing the movie I robot where spoiler alert, but you've had like 20 years to watch it. Like the big AI that's controlling everything, things it needs to protect all humans against like what they would want, but in their best interest. We also do have a benchmark right here for missile line behavior. which does include the mythos preview which is kind of exciting. Following that, there are some additional things launching today and I know I don't like to verbosely read through these but some of this is actually interesting because the dynamic workflow's portion right here is essentially in like... allowing Cloud Code to embark on long running tasks. It talks about parallel subagents. It seems pretty cool, but I would also be concerned about how much usage a lot meant this would drain. And if using this just through the API, this could probably get expensive rather quickly. Additionally to that, they do have some prominent mentions of effort control. To the point where if we scroll down a bit more, we actually see a- pretty big note on effort. Now when 4.7 came out, Opus 4.7, they introduced the adaptive thinking. I did not have that enabled when testing that model in the video that I did. And I found my experience with that model seemed to be totally different than what I saw a lot of folks reporting online. It was fantastic. It did the best subway FPS. I know that's only one, like tiny measure of a model's capability, but I found that a lot of people saying like this model's terrible had adaptive thinking enabled. So I didn't have it enabled and it did really well. So I think this is kind of a response to a lot of folks who are seemingly upset and experienced perceived or actual degraded performance with the adaptive thinking. Then this is... probably one of the more exciting parts. Well, first and foremost, we see users will find open 4.8 to be a modest but tangible improvement. That's like some pretty unexciting lingo to have a new model announcement, I would say. Even if that's the case, I mean, I don't know. There. They're doing well, they're confident. So I guess it shows some level of confidence to say that. But the exciting part of this is not only that, but we plan to release a new class of model with even higher intelligence than Opus. So for now, Opus is the state of the art publicly available and thropic model. But they talk about Project Glasswing, Claude Mithos Preview, and basically working to ensure that it has proper safeguards to be available to all customers in the coming weeks. So this Sentence right here is actually pretty exciting and would imagine a lot of folks will be talking about this. Finally the thing we have down here is just talking about pricing. Where it is the same price as Opus 4.7, $5 per million input tokens and 25 per million output tokens. For Fast Mode it is $10 per million in and $50 per million out. So to begin we're going to start with the Tried Intrubed Browser OS test which of course needs to make a browser OS with two functional 3D games, one of which is a

  2. 4:39 , obre el vídeo en una pestanya nova

    GTA clone, the ability to change wallpaper and a special feature it decides on an implement on its own. So it did try to open it in the artifact window. Fortunately I caught it in time so there were no spoilers. 910 lines of code. Alright, that's not a lot. I thought there was going to be more, but it does give us this handy opening Google Chrome button. So let's take a look at our Opus 4.8 browser OS. Alright, so I love synthwave style and like outrun and this has totally bothered me up to point where like we don't even need to go through this. It's the best one ever. Let's move on to the next. No. Alright, so this is this is awesome first and foremost so far. Does it have a right click? Doesn't have the right click but that's okay. We do have the clock in the correct time in my locale in terms of what this laptop thinks the time is. up on the top right, which is good. Neon OS. It's pretty prominently talking about this press command K for spotlight. Oh no. awful. I mean, it's cool. It's good. I just ate that feature on Mac, so it's included some of my, these favorite features. We can change the warp. Starfield. Sick. Okay. Let's go back command K. I mean, obviously the outrun one, the synth wave is going to be, okay, that's cool. A blue matrix or plasma. It's like disgusting lava lamp. I like lava lamps, but that color scheme was just nasty. Alright, let's check. Let's just all start with our GTA game. Drop into the city. Can I expand this or resize it? No, I can't. Okay. I'm not even a continued testing yet. We're gonna have it fix the issue that is causing the GTA clone not to work, and then we'll move on and proceed with the test. That was significantly less time and like everything than I thought. Okay. Let's check auto city now. Oh, sick. Alright. Oh wow. It genuinely kept the F to enter exit car. Oh yeah, look at that. Oh, the big problem here is that it spawned to the car inside the building so we were stuck there. But these car models are actually pretty good. We have people models. Can I do anything like malicious? Alright, let's see if we can borrow this vehicle. Good. Floor it. Oh, yep. Alright. Hidden run. Plus 25. There are collision, there are mesh colliders with the police. Let's just get out on foot. They can't get us there. Exactly. Later, can we steal the police car? Okay. That's not something I've seen so far. All right, let's just get back into this car. Really, what's letting this down is the inability to actually fall screen this. So it's so tiny in terms of the, like, maybe I could just zoom in on it and net it in. This has one of the better car models I've seen. And the people are little caps. So, okay. I have to remind myself continuously to not spend too much time in the GTA result. In the browser result, moving on we have orbital which is a additional 3D game. Okay, it's like one of these space shooter ones which like every model includes as the second 3D game. I think once I've seen a non-space shooter second game included and I've done this test quite a while now. Can we destroy any of these things? Alright, well I didn't see any asteroids get smoked but that's okay. Next up, I don't want to go into settings. I'd rather go into paint. Good. And we do have a safe PNG option. Even the color palette it's chosen to use here is like Synthwave. If beautiful. Oh, okay. Let's see if it saves it with a transparent background. It did in some ways, but then the that was left over. So interesting. All right, uh, calculator. Good. This thing, the adherence to Synthwave style, which again is not at all anywhere in that prompt, is curious because the Google model that came out was like that. What is this 435? Gemini 3.5 flash was like that as well. Very odd behavior. No pad, okay. auto saves to this browser. Nothing special here, but we do have a word and character count, so that's good. Let's make sure the word count works. Okay, good. All right, and we have our special live wallpaper. So those are inevitably the special feature, which I would be quite satisfied with as the Synthwave rolling one is pretty awesome. Accent color. And it does work, and it is like Synthwave style color. So we can even choose a solid color background. That's cool. And then upload image. Let's try it. Oh, I don't think this system doesn't have any like fun stuff on it. Cool.

  3. 9:18 , obre el vídeo en una pestanya nova

    this little device that I'm working on building. Don't worry about it. But it does look good as a background, which boads well for how it would look just in general. Alright, and then we have about, I'm going to go back to the synthwave. And about I think was what was opened when we actually had this thing. Okay, overall, I mean, I like it a lot. I love the synthwave. I was a little disappointed that one. We can't actually resize the windows, and there's no button to full screen them. And two, there were just like, There was one issue where GTA didn't work off the bat immediately, but you know, it is what it is, but there's no right click. So, I'm not bad, and I love the Synthwave style, which I did not at all ask for, but it totally leaned into it and like. Next up, I've swapped into the co-work portion of the Cloud Desktop application. This is going to be called the mid-roll ad test. So, we have given it a bunch of assets as well as a prompt containing a 60-second voice script. And the task here is to create an animated SVG for a little presentation with the script as well for the sponsor of today's video, oxylab.io. So, we'll see what happens here. And then this will also give us a look into its ability to create some video animation just with SVG and... some bundled included assets. And as it says here as well, this is a complex 60 second animation. And something I do believe these models have not yet really been properly utilized for. Is there ability to actually create some pretty detailed SVG animated videos? Something that like these can actually make marketing assets and things like this for Instagram or Google ads or something like that. And it's actually kind of cool watching some of the ways models go about doing that. Oh, aren't all right. All right. Hold on. Spoiler alert. All right. And we have our results. So we can now open it. in Google Chrome and see how it did. If you're building AI agents, getting real-time data from e-commerce, real estate, or search is a massive bottleneck. That's where the OxyLab's web scraper API comes in. It gets real-time data at scale from almost any public website using over 175 million residential proxies across 195 countries. I drop it straight into my NADN workflows. It handles JavaScript, heavy sites, and scales instantly in the cloud. Oxylabs just also launched a headless browser for AI agents, plus a fast search API that feeds organic search results into your AI pipelines. I did do something with NADN with oxylabs. It was really cool. It was like a drop shipping detector. go to oxylabs.io slash Bjion for a free trial. Up to 2000 web scraping results free, no credit card required. Use code Bjion in all caps for 20% off all oxylabs plans. Okay. I would say this didn't a decent job. Some things I liked were the animations were very smooth and again this goes back to what I may have mentioned previously where these models can actually be used to make animations like this with SVGs. They're very good at throwing edits together and things like this. Stuff like this can actually be used then for like marketing purposes as we see right here. But beyond that, like for someone who is running a Google ad campaign and needed assets or video animation, the models are actually quite apt at doing that with just SVG. So it's pretty cool to see and we see like the animations and movement and everything looks pretty good So I was happy with this it was very smooth the graphics looked nice It did use the color scheme as properly denoted in the prompt and overall this was pretty smooth and Acceptable result from cloud. Next up is going to be the beautiful static subway scene test. But this is now intertwined with the subway FPS game result as well. So basically it will create a pretty static subway scene and then it will immediately use that beautiful scene to turn it into an FPS with some humanoid slash zombieoid enemies. Also, I know zombieoid isn't a word. I just always use that in this prompt. I don't know why, but I like to. Alright, let's check out our subway FPS slash beautiful subway station result line six at derelict. subway station lit by falling fluorescence in the amber glow of service lamps. Something is moving in the tunnels. Survive the platform. Oh, yep. Yep. Claw's always going to tease Subway. It does it have, it doesn't show the gun, the gun bullet. It doesn't show the bullet effect in the wall, which is so frustrating because Opus 4.7 did, I think, and it was like the best result. Actually, though, look at this. This is a flickering fluorescent lighting. Park view. Oh look at the close up there of like the on the rivets actually in the metal that's not something I normally see in terms of the detail. I Okay, oh look at the glowing eyes there. Oh, okay, oh, muzzle flash looks good I'm still so frustrated though that the ammo doesn't actually affect the environment which is like a that's almost like a downgrade from what I previously seen but this is still I mean

  4. 13:57 , obre el vídeo en una pestanya nova

    This is very very very good. It runs smooth. It's like space is to jump very good. Shift is to sprint. Oh yeah, it is. Jump sprint. Okay, you got try again. Something I want to note that I actually find interesting is it has a toggle for flickering lighting. Sometimes that may not be good for someone. You might have like a medical issue or something that prevents you from being able to play games with that. So that's interesting attention to detail there. Now one of the biggest issues I'm noticing. The brightness slider is accessible if we just press escape. The cursor goes back. So now we can just quickly explore the actual environment. Again, the actual metal rivets shown here in these poles. I don't ever think I've seen detail such as that. But something that is a bit frustrating is, again, probably for the six time, the ammo doesn't actually take effect in the environment, which is a little frustrating. But this is still like this is very good. We even have a bit of camera shake like actually when we're firing the weapon model looks pretty good Let's just make sure the flicker toggle works. Yeah, it does cool So I've swapped into the cloud code portion of the desktop app I have dropped the effort down from max just to extra in here and we're giving this the self-contained C++ skateboarding test with the light 90s California boardwalk is that I alright, I've to say this is frustrating because it's just like showing me this bike jumping icon here and it just says almost on thinking but it gives me no indication that it's actually doing something versus having encountered some form of issue. So I would say this is rather frustrating extremely so like what are you doing? Okay, so it thought for five minutes with like genuinely no indication that it was actually doing something versus like frozen. You know what it's fine. It's going now. I even had sent this follow up is something happening because it was just showing me this weird orange jumping thing and saying almost done thinking like Doing not doing but now it's doing something so that's good. Alright the game has now been completed. It's going to compile it to verify that there are no compilation errors and then hopefully soon we'll be able to play with it. This is taking a little bit of time as we can see in the total clock right here. It's been right around 10 minutes. Alright and it gives us the specific command in order to actually play this 22 simple figures for NPC's. Skate game. See, this is good. Look at the walking in the NPC's. Is the word I was looking for here. This is very, very boardwalked like, okay, cool. The water looks good. We have nice palm trees. We have store fronts. I'll say this is probably one of the more organized results I've seen just in terms of everything kind of being laid out in that boardwalk style. Good tricks look good. Oh, that was a sick trick too. There are no mesh colliders on the NPT and I wouldn't want there to be because that would just be like two unplayable. Maybe. Yep. All right. Now let's do a sick flip off of this. Oh, back onto the rail. Right. I thought the board landed up. No, not it's fine. All right. And our skater has a hat on as well. Okay. It's not infinite, but it's fairly lengthy. Oh, what was that over there? I almost did it draw the stores. backwards because I swear I saw an awning like in the corner look but I can't get over there to check because it has boundaries that we can't cross up still though this is it's quite put together it may not be super visually exciting but I will say this is one of the more just like non-gletchy results I've seen I guess could be said and this is a more difficult test this is a test that very much separates models capabilities When sometimes like it may be a little harder or more vague to test just based like a browser OS because a lot of models can do a good job on that now not so much with this test so I do wonder though if the stores are inverted because it looks like there's an awning over there so I think these are oriented in the wrong direction but unfortunately we can't really see but aside from that would say this is Indeed, satisfactory. For this next test, I'm going to go to Ultimate Play.com, which is my own website where you can upload and play single file HTML generated games. And I want to find a game that was made with GPT55 because this will be a pretty good test. So this was the GPT55 result. This little beat them up game for Jerry's apartment in Seinfeld. But this started with two prompts. So the first prompt was just to create an accurate model of Jerry's apartment from Seinfeld. So I'm going to copy this first prompt right here. And we'll try to do a side by side.

  5. 18:36 , obre el vídeo en una pestanya nova

    I basically replication test of this, but with Opus 4.8 to see how they stack up. And I do have it still on max to be honest with you if I run out of thinking tokens or whatever a lot of them, I'll just up the subscription for testing purposes. Subscribe. It's properly searching and trying to find specific information. All right, so here is our Jerry's apartment result. Okay, click to enter. This is fine because it It's frustrating because this looks nothing like the layout of his apartment Where this to be judged individually it would be an acceptable result just based off of like a 3.js room being drawn But I'm not I'll say the bike Layout on the wall is probably one of the more bicycle like shapes I've seen in doing this test make of that which you will but overall we have the Mac we have these things, but this is not I expected better to be completely transparent. Rolling stone. Okay. That's good. The elements when looked at individually are acceptable, but if we're doing a side by side test with the 55 result here, I think I don't necessarily need to say much based off of what we saw online there. Regardless though, because the follow up for the specific test is to have it create. Okay. See is space up and down. I just want to make sure like I. Okay. So that's actually cool. That allows us to change the viewing angle that we actually have, which is not something I've seen before. the next portion of this when based off of this is to now turn this into a low poly beat them up game with fun arcade effects and animations. So we'll see what happens now keeping in mind that the apartment is not necessarily as adhering to the real life one as I would have hoped it would be. I do also want to check on usage. Okay, so 30% used in the current session. So gone up from 18 for the specific things we've done since my first look at this. So far, I like what I see better on invite a guest keep banging uninvited guests keep barging into Jerry's apartment close talkers low talkers soup cooks and male carriers funny and lore associated with the show W a s d j and k and l all right so L oh, okay, this is good. It's so frustrating because the actual like animations humanoid wise here. Okay, that's cool. That guy just got the Renity now super. Okay, it lags a little bit and Like we see a very pretty city effect outside the window which is glitching a bit but that role effect is this is a this is a good beat-up game I will say it's not It's tough to judge but I do like it. It's fun and again this was more of like almost like a versus GPT55 test the movement is cool I would kind of like to have different camera angles that we could see Oh, oh, that was just That like freeze frame there when we got KO'd. Looks painful. It's tough to, I mean, this was done through Codex as well, which is more natively like a coding slash front end thing. But I do think some of the fighting logic was perhaps more fun in the opus one, but I don't know. That's definitely going to be an ask the audience one. The movement that was actually, there's bending knees which is interesting to see, and it's tough to say. Next up we're going to be doing a classic test in the form of the 3D Flight Combat Simulator game. It's always fun to play the results of these and see how well they do with the plane models. So it's written the file, but it seemingly like just frozen up again. So I did copy the contents of this specific file, but sometimes it writes it and then it's like, oh, I found one thing that I'd like to fix. Though I do have half a mind to just manually paste those into an HTML file right now because I'm not quite sure what's going on. But we have something so. So here's the early plane result. A, Stominian. All right, these models look good too. It's hard to fly. Sometimes that happens when they actually have the cursor control the plane. But this is overall, again, it's well done, especially because of the plane models. Now obviously one of the big issues we're noticing right here is this frame rate is abysmal. Oh, when the ammunition hits the ground, we actually get FX, I think.

  6. 23:15 , obre el vídeo en una pestanya nova

    It's actually putting like a box around the enemy. All right, this is it's frustrating because I gotta be honest with you I keep waiting for this model to like wow me and it's not happened yet and I'm a little upset like we're seeing hints of it like the way the ammo tracer's actually hit the ground and stuff like that I think that may be attributing to the large large amount of lag we're noticing but that was really impressive Let me go back into the cloud app right here to see if it has done anything now and it's just frozen up right here, which is couldn't finish the response try again in a moment and then if I Yeah now it's gonna rewrite the whole thing again, which is gonna smoke my usage limits I'm honestly gonna try turning on the adaptive thinking right here I'm keeping it on max, but I'm almost wondering if something's wrong with that I'm gonna put it on the default effort and have adaptive thinking on for this next test and we're going to be giving it a front end design test for the Ravioli Rousseau. This is the Ravioli brand front end test that I have been playing with lately. Alright, let's take a look at our Ravioli Rousseau website. Okay, very good. We do have the SVG Ravioy's floating around. This is a very detailed prompt, which basically means that every model I test is with that's at least some are competent. Produces very, very similar looking results. And when we click in the background, we do spawn a bunch of Ravioli and related Ravioy things. Let's check our simmer richness slow simmer tomatoes. Okay, and we have some movement on these card effects. Now the big test is obviously the jar being drawn with SVG. Okay, three things we refuse to rush the long simmer roasted garlic gold garden basil late where is the jar? Very, I've seen better. It's there. It's floating, but I have seen better. Your pasta has been waiting taste the sauce and we have a little tiny photo shifting glass. I've seen I mean, this hasn't been run very much, but all of the models produce very, very similar things. So, this is in line with what is normally seen with this result. So, I've put it back to Max effort, but I do have adaptive thinking still enabled now, and we're going to test the 3D printer simulation test. So, the response froze when I had it on maximum effort, and it was thinking for quite a long time as we saw, and then it... Stop working and then I got some connection errors here. So I've swapped it back to high with adaptive thinking enabled and we're running the 3D printer sim again in the hopes that we get something before it glitches out or freezes up. So I'm not I'm not enjoying this right now to be completely honest with you. I'm I'm not happy with this model. I'm I'm actually watching my Opus 4.7 video back here just to make sure I'm not the one hallucinating and every result I've tested With this that I tested with 4.7 the 4.7 results so far has been more impressive. Alright, let's check out our 3D printer simulation result. This was on high thinking mode with adaptive thinking on. Alright, so hypothetically we now have our fixed version. Good. This is very good and this is kind of what I was hoping to see more of with this model. Keep in mind this was done on high mode with adaptive thinking enabled. I did try it on Mike Max thinking, but we encountered an issue after many, many minutes of thinking. So we do have a small connected to the nozzle. The nozzle looks good. We have an LCD screen. This is a Core XY style printer. Let's just go ahead and take a look at the print. Oh my, with the heck. What sort of squares this I will say? It's... So artistic. So 3D printers are very fickle creatures. And if you have had the dysflesia of having to tune them and play with them, you may notice that this is actually a fairly realistic result of what a very very misconfigured 3D printer would produce when trying to print a square. Now I find this hilarious, but the actual printing is going line by line and the infill is more or less like proper is this spool at all spinning. It is good. You can see that good and the gantry the movement is smooth and everything like that It's just quite perplexing this square. I'm gonna maybe just Okay, we're 80% let's just make sure we get to 100. Okay, let's try the circle I will slow the time scale down Okay, good. So the circle is a proper circle The square was also somewhat of a partial circle So, okay, did we just change film and color midway? No. It's showing us a different color for the current layer. It's on, all right, and this is a speedy printer. Let's just, let's just put it to max speed. All right, not bad. And then finally, we'll try a triangle. We'll see if this is, nope, it looks like a Hershey's case. Very, very,

  7. 27:54 , obre el vídeo en una pestanya nova

    Interesting, although check this out, the spool color is actually changing depending on the shape we're printing so it changes depending on C. So it was blue for the circle yellow for the squiracle and blue for the... Alright, interesting, you know, probably done more or less the shape was good, the nozzle material looked good, the actual drawing there, the spool spun, the movement was good, very, very interesting. I'm going back to this flight combat simulator thing because it was just like fix this 3fps looking trash now. Alright, we see a preview here and that actually happened fairly quickly. So the plane models do look much better. Let's just try the Falcon X. Okay, it's it's like a first person view and we're going backwards. Okay, some oh no, this is this is not okay. It's not okay. Right, there's supposed to be a two second black screen when we crash not an infinite one. Let's check the war Hawk at minimum. We can see the actual plane models there at least in the thing and it did try to simplify it to its benefit. So it was focusing on performance, which is why it would have detracted from the actual graphics or aesthetics here at least the plane models look better and it runs quickly. Alright, let's check my usage right now 63% used and I'm worried I may tap my usage limit. And if I do I will just by the more expensive plan, because I want to give this thing a chance to really properly redeem itself so I've prepared a difficult test. Alright, so I've genuinely been... disappointed with this so far, especially after looking side by side at the Opus 4.7 results with what I've seen so far in this test. Everything here has been worse than what was experienced with Opus 4.7. I want to give this thing a really difficult test and I have put it on fast mode Opus 4.8 max and we are using the code version or code window of this desktop app. I will inevitably hit some usage limit here and I'll either just make it use API credits or something else, but this task is going to be difficult. I've given it this photo this arcade machine and told it it needs to make a website with an accurate 3D model of that using something like 3JS. We can close that. Additionally to that I've given it this sprite sheet that was generated with GPT image 2 of 10 different cars. So it's... Instructed to turn this arcade machine into a cool 3D website and then make a game using these sprites put it on this digital screen on this website and then have the user be able to play this using the WASD keys. So it has access to these assets and it is going to work here in Fast mode just to give this thing a proper chance to showcase its capability because truthfully I've been really quite let down so far. So um more than happy to spend a lot longer testing it to give it like proper chances to shine, I guess, could be said. Now what it's doing right here is interesting because it's just working around trying to pull the black cars from the image without the background also being in there. So that's what it's talking about separatable with threshold 12. And part of the reason this is a difficult test is because the models will try to lean on knowledge of like dealing with sprite sheets that they may have. And then that will fall part quickly when they try it on this AI generated non-traditional sprite sheet. So then they have to figure out how to... work around that to actually still get these as sprites. So it just makes for a more complicated task of having to do this. And don't forget the whole part of this is do then create a 3JS 3D model of the arcade cabinet, build this game using these sprites and then place it on that and put it in a website to make it look good. So this is a proper intricate test that I do want to give this model because so far it's been so frustrating to deal with. Okay, so That's acceptable. We see the sprites work well and things like this. Now keep in mind this is just a test of the game. Good. The sprite movement does if we hit cars we do get like we slow down and this is okay. Now if you recall I had told it I want full 3D not pseudo 3D and it's also supposed to actually build the arcade cabinet in 3D and then render this on the screen of that. So

  8. 32:33 , obre el vídeo en una pestanya nova

    This is just one piece of the puzzle still, not fully 100% there yet, but mostly what happens. All right, we should. That is kind of sick. It doesn't 100% look one to one now. The machine looks, but that's a... That's a pretty sick UI preview and things. The steering wheel is going through the desk. I would really hope that it will be able to pull that out of the screen shot that it takes here to look. And then remember the whole goal of this is to actually transpose the game onto the arcade cabinet right here. Light reflecting off the scene and things like that. This is good and this is more in line with what I'd hope to see Please tell me that you understand the cabinets form and detail are lost in black and I want to check the layout against the reference Interesting because the cabinet doesn't necessarily look like this in real life Based off the picture and what we saw it like this is attached to it and it actually comes out from the side here and all this looks more or less correct Just in this like specific steering wheel area. Okay It changed it a bit. The joystick became a different shape from what we had there which probably a bit more in line with what I would have wanted. The steering wheel is higher and no longer actually hits the ground which is good to see. And if we actually zoom in here a bit, we should be able to see some layer lines for 3D printing. Yeah, I think they're there. I don't know to be honest with you. I can't tell, but it does have visible bolts and things like that. The steering wheels You know, let me check in your head on angle matching reference to Judge proportion. Okay, this is taking forever. I just I did this to myself though because I want to see this model like Try because so far I've not been wow all right. This thing's like going crazy about like fine tuning the arcade cabinet. I'm stopping it and saying the arcade cabinet is fine All right, so I've given it this follow up here and we'll see what happens Alright, so it seems like they 3D edit for this game is going to be visible to us soon. Okay, I'll be racer press W Fantastic good. That is good. This is what I want it. This is very good. It utilized the sprites properly to make it 3D more or less. The road has some oddities and things like this, but it should properly Photoshop. Photoshop. You should take a screenshot of this and then understand that it has a few things that needs to fix. Very good. And this is so far, probably taking an hour for this specific test. I don't necessarily spend this much time on one test, one frontier model drops, and I want to test it. But because I was in am disappointed with how this model has performed so far, I really wanted to do something like, this is genuinely 3D now. Yes, it is. I wanted to do something very, very intricate and See what we get. It's gonna like maybe move the camera here because okay, see now this see it actually is Functional to this is Absolutely very cool that weird like glow of the light reflecting on the screen is kind of detracting from the gameplay, but the fact that I mean, this is just really, really kind of neat. Does the joystick actually properly respond to forward backward movement? Yep, it does. And the steering wheel moves as well. And the thing we want to do is just test it actually in Chrome. What did it call it on the eraser? Okay, so this is deeply, deeply impressive. And this is exactly what I wanted. So the whole thing of this test was it had to create the 3D model of this arcade cabinet off of a reference photo. It wasn't fully one to one, but that's okay. Then the entire other thing was to take the unstandardized sprite sheet that we gave in. And then have it create a 3D racing game using the assets from that. It is difficult to take these sprites and make something that looks 3D. And then the whole thing was to place it on this arcade cabinet. So the user could actually play it just using WASD. And that is exactly what has it. heard right here. This was a very difficult test that probably took about an hour and a half for this to finish end to end. But the whole thing was I wanted to try something else because I had been rather disappointed up to this point with some of the previous things we had seen as a result from this model. So this is promising. And I mean, like, look at that. That's like sick even from here, like from this view as well. So if anyone is interested in actually seeing this, If you go to altimplay.com, which is my website where you can upload play and upload ad generated games, I will post this on here for anyone interested to actually be able to play. For the final test, I'm going to be doing a drum kit simulation test, which includes the auto play feature, so we'll have four distinct tracks that we will be able to auto play.

  9. 37:12 , obre el vídeo en una pestanya nova

    is just using the Claude AI web chat because I wanted to just run this at the same time that we were doing the arcade simulation, which was very, very impressive. This is on max and if we go into effort, we see is adaptive thinking on okay it is, but I just have it on max as well. So we'll get this and then we'll summarize. All right, so it reached its max length for this message for the drum kitsim. While it was just still thinking, you've used 90% of your session limit. Yeah, that's fine. This is the last test I'll be doing with this model. Here's our drum kit simulation so far this looks like it may be very good and it does seem very good. Okay good well Everything is laid out more or less correctly actually really the only slight issue I see here is like this should be pointed the other way, but that's actually well done. I'm just testing it here See those animations are very very good look at the actual curvature to the symbols I don't know that I've ever really seen that before so that's good attention to detail that's well done Fill Holland's test is good Let's check the auto play feature so first and foremost we have rock that slaps good the sequinsers good or the sequence and we can change the tempo wait for it That's good. All right, let's try funk Alright, let's try hip hop 86 BPM. That's very slow. Yeah. Oh, testing this model for five hours makes me sad and sour. It's not just this too slow. Then jazz. So that brings us to the conclusion here. And truthfully, I'm very very, very, I can't think of the word, conflicted. Because This result was nice. The Omni-Racer result here was absolutely sick. And as a matter of fact, this did actually have sound to it. I just had Nikleck did to pick up on that earlier when we tested it. So this result and the drum kit result were fantastic. The problem is when we go to the other things... I really found myself rather not impressed comparatively to the previous Opus 4.7. So for example, here is the specific GTA result from within the browser OS test. We can go, we can enter cars, and you can knock over pedestrians and things like this. This was in the style of like the synth wave that it made the browser OS for some reason, which was interesting, but you know it did work and the pedestrian just disappears there when you hit them. So if we compare that to the Opus 4.7 result of the same exact test the pedestrians actually fell down and stayed down when you hit them. There was moving traffic autonomously around the map and this overall was just a better result. Not even mentioning the fact that the Opus 4.8 result actually had a specific error prior to this working in the first try. So this is Opus 4.7 for that same result we just looked at which is like one thing that made me a little like, oh, okay, I'm not 100% impressed right now. Next up we had the subway stations on the test, which was pretty good. The muzzle flash effect and everything like that was impressive. However, this is another one of those that, although this was good, the Opus 4.7 result that I was testing, or had tested when that came out was arguably better and I'm going to once again do a side-by-side dish, okay, so that. So this is the Opus 4.7 subway station result. Yes, it is a little darker, but there's a train there. It has a smaller, less emphasized muzzle flash. But the one thing here that was really quite impressive was when we actually shoot the environment, the bullet holes actually appear right there. And you can see from my face in that video, or maybe you can't, because my face will be in the same location that faces there.

  10. 41:51 , obre el vídeo en una pestanya nova

    I was very impressed with this. Now the argument could be made that this was a two-part test where first and foremost we just had to create the static subway station and then in a follow-up prompt in that same conversational thread we then had it turn it into an FPS as opposed to just combining those two prompts for the OBS 4.8 result but this was just such a significantly improved result that this was the second thing that kind of made me like, I'm not really impressed right now. So here was the Opus 4.8 Jerry's apartment result, where, okay, everything looks alright. It looks nothing like Jerry's apartment from Seinfeld, but it's, you know, it's there. It's like somewhat together. Now again, if we go and actually look at Opus 4.7 to see the test result in Jerry's apartment, we can see right here. I think that this is a better result than what we received from 4.8 in terms of the apartment model. It's more properly laid out in the style, I guess, of what Jerry's apartment would actually look like. The kitchens more in that area. I honestly think there's more finite detail in here as well as opposed to, oh, but I had given it a photo reference there. Okay, so that is perhaps a knock of B. John. I will say that. Versus just giving it the prompt here. So I'm glad I caught myself on that one. So that changes things a bit. We also had the C++ skateboard game. Now this one was not bad. It was pretty good. Everything was clean. It was properly boardwalk style. The tricks looked good. The water looked good. But again, if we now go back to the Opus 4.7 test video, we're going to see that the result that that made at the like N1 was significantly more impressive than what we received with Opus 4.8. So if this ever loads, This is a more impressive result, just the way the water's drawn in the storefronts are detailed. The actual effects and everything like that was... just way better than what we got with Opus 4.8. Now the one caveat to this is that it did initially do AC plus plus test and it actually just made it in 2D. This is the Opus 4.7 test video. So this was a later follow-up test where I took the result that had created in 2D and said make it 3D versus just giving it the one prompt saying 3D already in the prompt for what 4.8 gave but still this was like better in A lot of ways although the people moving around I guess would not be They were better with Opus 4.8 so I just wanted to give some level of actual side-by-side showcase because I was not really blown away with this model to the point where I actually made a post on X saying that and I never really say negative things about the models But this one just out the gate did not really impress me very much And later it may have kind of fixed itself. Oh wow, this is actually running quickly now, our sky dominion game. And the ammo tracer's hitting the ground actually showing visible explosions in this were sick. Look at that. I've not seen that with any of the results before. The Reveal-Eroso website was the bottle was disappointing. I saw a better bottle from Q1 3.7 max review as well as Gemini 2.5 flash. The 3D printer sim was not working well on first try. And then later on that... Square was just breath like Omni-Racer. This one right here is hands down the more most impressive result in this test and perhaps one of the more impressive results I've shown in a YouTube video but this took a long time in the code mode with the fast mode enabled and our drum kit designer was pretty good as well. So Overall it's very hard for me to judge this, but I was not initially blown away with it as I would have perhaps expected to be and I'm very interested to see what folks think of this model as they themselves use it beyond just watching this video and seeing my experience with it. But this video is probably going to have been significantly longer than normal and I wanted to really test it with something more intricate because I was initially kind of like, uh, and I want to back that stuff up. So Basically, that's what leads us here, and this is where we'll close out. I will put this on, UltimatePlay.com, so anyone interested can go and play this themselves, because this is a sick result. So that's going to conclude our first look in test of Opus 4.8. I will be very excited to see the mythos-style level models that will be released in the coming weeks, as well as perhaps GPT-56 and Gemini 3.5 Pro, as well as Mini Mac, 7-3, and stuff like that. So with that, Thank you for watching. If you have any questions, please feel free to leave them in the comments.