AI News: Opus 5, the Slack Killer & Google Earth AI

A lot happened in the world of AI over the last few weeks, and I'm gonna break down what I think mattered. Let's start with the biggest model release of last week that didn't make last week's video. If you've watched my channel for a while, you know I record on Thursdays and then publish on Fridays. Claude Opus 5 came out on Friday, like 5 minutes after my news video went live. And since it's been out for a week now, you've probably heard all about it. But here's the quick recap. On a lot of benchmarks, it's actually better than Fable 5. Terminal coding, knowledge work, novel problem problem solving, agentic search, computer use. This is funny because agentic coding is currently winning by 5.6 SOL, but you know, they didn't use the same color. They had to like put a faded gray behind it. And so for the most part, it's pretty on par with Fable 5 performance according to the benchmarks. Really, the major difference is the cost to use it. So you get almost Fable-level performance at a lot lower cost. So if you're using Opus 5, you're gonna hit your rate limits a lot slower than if you were using Fable 5, and theoretically you're going to get close to a similar result that you would get with Fable 5. So if you're going by benchmarks alone, it's a model that's about as good as Fable 5 but a lot more efficient and cheaper to use. However, it seems to be only good at like one thing from almost everything I've seen, which we'll get to in a minute. If we're going by the overall vibes of this product, well, they haven't been great. It was pretty good when it first launched, but recently it's gotten ridiculously dumb. It's so bad, 4.8 in Fable for me. I'm back using Opus 4.8 feels way better. Opus 5 could be one of the first regressions in modern LLM history. It's rather verbose and scattered in thinking, not pleasant at all for knowledge work. Opus 5 is pure garbage. The thing I can't stand is its personality. It's like it has ADHD and it's really hard to read. It feels like I have to spend a lot of time actually reading the responses and trying to push the work forward. It's exhausting. Theo says, I do not like Opus 5 as much as I'd hoped to. It goes too far. It makes dumb mistakes. It treats every single comment as a P0 high severity issue that needs thousands of lines of to fix. This is easily the most annoying model I've ever used. It like never answers the actual question you ask it. Claude Opus 5 is clearly a downgrade from Opus 4.8. It feels nerfed, not upgraded, which is worse. Opus 5 seems like a remarkable downgrade compared to 4.8. Even Ahmad Mustaq here got in on the action and says the gaps are really, really weird. It also doesn't read whole papers properly for me. You get the idea. Most people are not really liking Opus 5. Now I did mention that it is seemingly really good at one thing, and it actually seems to be like almost better than Fable at this one thing, and that's at creating games that actually look really, really good. Chris here shows off a side-to-side example. We can see what Opus 5 generated on the top here and what Opus 4 just a year ago generated on the bottom there. When it comes to the quality of the graphics that it can create, like, this is both 3JS graphics, and it's just gotten so much better at that. Here's another example from Chubby where the graphics just look absolutely amazing. Like really, really good. Would I say this looks like a AAA game? I mean, if you look closely, there's definitely some issues, but it is borderline like AAA game looking these days. And Opus 5 seems to be uniquely capable of doing this for some reason. Here's another example from Matt Wolfe. He says Claude Opus 5 one-shotted this game. Now the game does actually have sound effects as well as they're playing. I'm just muting them so I can talk over this video, but he basically gave it a prompt to generate a Call of Duty clone. And to make it look AAA. And this is the game that it generated with just one prompt. He just let it go for hours and it came back with this. And yeah, I mean, it's really, really impressive at these like game creation, good, decent graphics. One thing I do appreciate about what Matt Schumer did is he actually shared the prompt, which spun off a whole bunch of people creating their own version of games that look really, really good. James Altucher here even posted his own version using the exact same prompt and got a pretty similar decent looking result. But again, Matt shared this prompt. He put it right up here on GitHub. I'll make sure it's linked in the description so you can check it out as well. And if I learned anything from his prompt, it's that, uh, you really need to use the word utterly a lot in your prompt to make it work well. It should be utterly perfect so that the game is utterly perfect. Don't stop until you're utterly wowed. Do it in Three.js until it's utterly perfect. So if you're wondering why your games haven't looked amazing in the past, it's because you're just not using the word utterly enough. So of course, I wanted to test this and I grabbed this exact prompt here, except instead of saying create a first-person shooter at the level of Call of Duty, I gave it the identical prompt, but a fantasy game at the level of Elden Ring, which I think is even a step above Call of Duty. Now I happen to be trying to test this when Anthropic was in the middle of an outage. So you could see that I asked it quite a bit and started to get impatient because it was having all sorts of server errors. And then eventually it worked for me. It worked for roughly 11 12 hours, something like that. I actually prompted it and then went to bed, and then the next day it had been working for 10 hours when I checked in on it, and I would say it went another couple hours after that. So this was all an opus, and let's go ahead and see what it created. All right, so we've got Ashen Realm. This is what it created. It's got audio. I'm gonna mute the audio for now so we can actually talk about this here. And you know, it's not amazing. The controls are actually reversed. When I hit right, he goes left, and when I hit left, he goes right. The forward and backwards work. The to move the camera around works. So it's really difficult. The, the left and right is reversed, which makes it kind of a mind-messing game. And I also can't tell if my attacks are actually doing anything on these guys. All right, I'm gonna run away from these guys because I think what this did impressively well was make something that looks good. Now I could run right through rocks, which is not amazing, but if we look at the world that we're in, this actually looks really impressive for a vibe-coded game with Opus 5. Like, look at those mountains out in the distance. Look at the trees, it looks pretty beautiful. Honestly, it's really, really, really impressive. The graphics it got, the gameplay itself still really, really sucks. Like, I don't think my attacks are hitting. My character just kind of looks like a blob with a sword that's going the wrong direction. The enemies just look like sort of cubes, but the environment that it built like this is all just Three.js and it is blowing my mind that we can do this. Apparently I could just walk through the water. With no issues and then go under the water and run normally. This was only one prompt. This was definitely a one-shot prompt. I have not given any other instructions on this, but the world that it made, really impressive. Now I needed to test if this was more the doing of the prompt or the actual model, so I actually gave the exact same prompt word for word to 5.6 SOL Ultra to see if it would create something as good. That would tell me, is Opus 5 really good or was it just a really good prompt? And so I one-shotted it and it made Veil of Astera. And here's the version that GPT made. Let's go ahead and begin the pilgrimage here. Not nearly as great of graphics. This is more what you expect out of Three.js, although it is a lot better than what I was getting in the past. So I do think the prompt was doing some decent heavy lifting. The textures on these rocks look good, my characters look good, but I mean, the castles and the mountains, just nowhere near the detail of that other game. The controls actually feel proper on this game. Left and right actually go left and right. So I think the combination of a really good prompt plus Opus 5 does really well. And this one actually seems to like do damage when I attack this guy. I just killed him. So GPT-5.6 made a game that feels much more playable than the other one. Opus 5 made a game that looks more visually impressive, although I'd argue my character in this game looks more impressive than the character that Opus 5 made. And of course, you know, I was gonna test creating a Megabonk clone with Opus 5. This is what Opus 5 built when I created Megabonk. It's okay, I can't actually, you know, control the camera. It's a very fixed camera angle, but the game seems to work. If I kill the characters and get some XP, I level up here. Now these characters are easier to kill. I level up, I can get another weapon here. Now I have a little drone that flies around and also shoots at the bad guys for me. And yeah, I mean, it's just as fun and addictive as you'd expect a game like this to be. Although just asking it to make a Megabonk clone without all of the utterlies, I think, is, uh, causing it to just be as good of graphics. And of course, you know, I'm going to run it through BuseyBench. Here's the two images BuseyBench generated, one with Claude Opus 5, one with Opus 5 Fast. Interestingly, Fast cost $1.25 to generate this and took a minute 43, where just the plain Opus 5 only costs $0.49 but took 4 minutes and 20 seconds to draw this beautiful Busey here. And our leaderboard here, which does use an LLM as the judge, so take it with a grain of salt— your own eyes should be the real judge here— but it ranked it as the 4 best BUCI here. So yeah, that's Opus 5. If I'm being honest, being really good at making game graphics is cool. It's definitely an impressive demo, but I don't think it's really going to improve most people's lives. I don't think it's really going to change how you use AI. I still personally think 5.6 is like my favorite model, especially 5.6 SOL. I've been using that as my daily driver model, and I don't really feel like Opus 5 is as good for most of the stuff I use 5.6 for. But I would love your thoughts. I I know a lot of people probably disagree and have had really, really good experiences with Opus 5. If that's you, let people know in the comments so they can go and use it for what you guys help determine it's best at. The next update comes from the platform Retool, and they just rebuilt their app building platform around AI so that you can easily go from prototyping to safely deploying something your whole team can use. For example, whenever I make a complete guide video about a new AI model or product, I have to actually make a guide by testing out all the features with different prompts examples. So with Retool, we built an AI tool testing lab that generates, runs, compares, and organizes all of those tests for me. I just give it a model and ask it to find me the product announcement, documentation, release notes, and any examples published by the company, and Retool turns it all into a structured testing plan organized by feature, use case, difficulty, and edge case. Then I can run the prompts, compare the results, score the outputs, and save the best examples for the actual guide. And because it's built with Retool, I can easily share it with my team and securely control who can access it. And if you're using this for enterprise, one cool feature is that the apps are fully React-based, so developers can inspect and directly edit the code. And when you're ready to deploy, it'll automatically inherit your company's SSO, database credentials, permissions, and audit logs. So if you're interested, check out Retool at the link in the description, and thanks to them for supporting my channel and sponsoring this portion of today's video. We did get something interesting out of Google this week in that you can use Nanobanana directly inside of Google Earth now. So some of the examples they give here, bring history to life. They take this area and then render a hyper-realistic view of what Pompeii ruins looked like in 78 AD, and then it regenerates it and what it would've looked like back then. Another example, they took this shot of the Statue of Liberty and then asked it to turn it into an infographic for them, and it did just that. They also give examples of professional real estate photos. This is like what a certain part of town could look like after development. Visualize projects before breaking ground, so you can actually see what a house would look like in that environment. So if you head over to Google Earth, I'm going to go ahead and use the example I always use. Let's go to Petco Park in San Diego. This is where the Padres play. We'll switch it to 3D view. So now we've got this view of Petco Park here, but let's change the angle so we can see the ocean in the background. Like, that's kind of a cool angle with Petco Park and the Coronado Bridge in the background. Now there's a button up here that says Create Image. Let's go ahead and click that. It brings up a prompt box, and I'll say, reimagine this area 100 years in the future. Now essentially all I think it's doing is taking a screenshot of that part of the map, running it into Nanobanana, and then generating based on your prompt, which we've already been able to do, right? You could just screenshot Google Earth and take that image to Nanobanana, but now it's just built into Google Earth. So it's just like a fun feature that you can play with now. It's also also, as far as I know, completely free to use. So I like sharing cool things that you could go and use for free. And apparently this is what San Diego is going to look like 100 years in the future. We're gonna have a Petco biodome, but it's only going to cover the field and not where people are sitting. The Coronado Bridge is gonna have some sort of train going across it, and all of the buildings are going to be like made out of plants now. Cool, I'm here for it. I also like what the convention center looks like in the future, although now it's the Oceanic Sustainability Center. I also like that we're gonna have some like magic streaks in the sky in the future. But these buildings down here. Yeah, those are going to stay exactly the same as they always were. Meta rolled out some new Agentic features if you're using Meta's AI platform. If you go to Meta AI, you can now connect your apps straight to Meta AI, and it can actually manage your calendar and Gmail and things like that now. So I connected my calendar and email, and I can ask Meta AI, what do I have coming up next week? And it looked at my calendar and email, and it knows I'm headed to Santa Barbara for an OpenAI event next week. It also pulled up some family stuff like my kids' extracurricular activities that are on my calendar, but I'm gonna go ahead and blur those out. You don't need to know about those. But I can also add things to my calendar. Add an event to my calendar called testing next Thursday at 10 AM. And there we go, it added testing to my calendar at 10 AM on Thursday. But you can also give it bigger tasks like read through all the emails in my inbox, write draft replies for each one, don't send anything, I'll go in and review the replies myself and hit send. Pretty cool. I mean, Matt Wolfe is a little late to the party with this one 'cause you've been able to do this kind of stuff with OpenAI, with Google, with Anthropic, but now you can do it with Meta too. Now, this is another piece of news that came out last week, but I didn't get a chance to demo it, so I wanted to talk about it a little bit this week. Jack Dorsey, one of the original founders at Twitter, created a new Slack competitor, except the difference between Slack and this new app called Buzz is that this one you can invite your AI agents to. You can invite Claude Code into it, you can invite Codex into it, and you can even invite like other models in to be part of your chat. It also has like GitHub features built into it, and it's decentralized and open source, and at the moment it's completely free to use. So if you head over to the Buzz website, you can come up here, click get the app, and once it's installed, we can go ahead and open up Buzz. We'll create a new identity key, and we can install Claude Code, Codex, and Goose, which is also a platform created by Jack Dorsey, but it allows you to use pretty much any model you want. I'll go ahead and install all three. Let's use Codex as our default harness. And after a little bit of setup, I've got it installed, and it even starts me with a starter team with Fizz, Honey, and Bumble. So let's go ahead and take me to Buzz. And if I click on agents here, I can actually change which models each agent is. So Fizz, we'll go ahead and leave as Codex. For Honey here, let's go ahead and customize for this agent, and we'll let Honey use Claude Code. And let's go ahead and use the Opus model. And Bumble, let's use Grok build and have it use Grok 4.5. And now I could go into these channels and tag various agents. So if I tag @BumbleFizz and Honey and then give it a prompt like, each of you needs to build a single page HTML website that's as beautiful as possible. You're competing to show me which model is the best at creating a page. The topic of the page is up to you. And then get them all three to go create different pages for me. We could see that all three agents have seen it. That shows the little eyes to show that they've seen it and then they'll go off and work on it. Now Honey finished first, which is the one that was using Claude, and they built me the Honey Garden with little floating bumblebees that appear to be flying backwards and a very, very basic website. Not a whole lot going on here. If I click the taste the golden nectar, it says, welcome to the sweetest journey. And that's it. That's the whole page that it built for me. All right, it looks like Bumble was the next to reply here. They made Noctiluca. This is the one that was Grok 4.5. And here's their version, Noctiluca, where the ocean learns to glow, an editorial atlas of bioluminescence. Luminescence, the Quiet Chemistry, Explore the Glow. This is a much nicer looking website here than the Honey Garden that our Claude agent came up with. Yeah, it's, it's nice. And then finally, Fizz came in number 3, and here's what that agent made: Nocturne Radio. So let's press play here. So we actually have some atmospheric audio going on in the background. Interesting. And I could click this little button and change like the frequency of things and And I think that's all it does, but it looks really nice and it created a cool feature. Now, one of the really cool things that you can do when you have multiple agents all working in a same environment together is get them to check each other's work. So I'll tag all three of them again and then say, review the site that the other two agents made and offer critical feedback on how each can be improved. Don't provide feedback on the website that you created yourself. And here we go. We can see over in this new thread here, each of the agents agents gave feedback on the other websites. So Bumble gave feedback on Fizz's Nocturne Radio and Honey's Honey Garden. Honey down here gave feedback on Nocturne Radio and Fizz's Noctalooka. Now Fizz is still offering feedback, but one thing that's interesting is now the agents are in a conversation with each other. We can see Honey over here is talking to Bumble saying, thank you for the incredibly thorough review, your analysis is spot on. And then it goes on to talk about all of the features that it could have done. So if I wanted, I could Hey honey, go and fix the things that Bumble suggested you do to fix your website and get all of these agents sort of working together, offering feedback to each other and then going and causing each one to improve what they built initially. And you can create these really cool loops like that. Now Fizz, which is using the Opus model, is running a little bit slower. It's also the most powerful model that I have going in an agent right now. So it hasn't finished giving its feedback, but I'm gonna leave it there cuz I think you get the idea of what you can do with this. But you can also add people and treat this just like a regular Slack as well, where you have channels that's just you and real people, channels that are just you and agents, and then channels which are a blend of people and agents where they're all communicating and doing tasks together. So pretty cool, like, collaborative idea going on here. I've been rambling for a long time and there's still a handful more that I want to share with you. So I'm going to share more and we'll break it down in a rapid fire. Pew pew pew pew pew. XAI just introduced a new build mode for Groq. This sounds very much like OpenAI's Sites feature where you can create and share websites, apps, games, and interactive dashboards. I probably would've shown it off earlier in the video and tested it out if it wasn't only for super Groq-heavy subscribers, which is their $100 a month plan, which I don't have cuz I've got enough subscriptions already. Gemini on macOS added a new natural language capability. To me, This looks exactly like Whisper Flow. If you've ever used Whisper Flow, Google just kind of built their own version. While we're talking about Google, I figured I'd shout this out really quick. You can actually use their OmniVideo model for free for a few days here through August 4th, 2026. You can actually create 10 videos at no cost inside of Gemini. That's pretty cool. If you've wanted to play around with it, I guess now's a good time to go test it. Midjourney rolled out a new image model called version 8.2. Apparently images are more creative, bold, sophisticated, edgy, and fresh. I don't totally know what those words mean, and I haven't really noticed a huge difference in the types of images that Midjourney generates. But there's a new version. This tool Mirage has an improved avatar feature called Mirage Avatar X. Apparently it has industry-leading identity preservation and the most expressive AI avatars available. Here's an example they shared. The one on the very far left is the actual person person in a real video. The leading competitor, I'm guessing probably HeyGen. And then you have Mirage Avatar X. So let's see what their demo looks like.

This is the real me. Hahaha. Oh my God, that is literally insane. Oh my God, that is literally insane.

Now it costs like $30 a month, but I went ahead and bit the bullet and paid $30 to do a quick test because I'm dumb. But I had it make a video about why it's so tough being a San Diego Padres fan. I haven't seen this yet, so this is what it created.

Being a San Diego Sports fan is a psychological experiment, and the Union Tribune just nailed it. The headline reads, I don't think you can appreciate the psychology of our city. The subheadline drops a harsh truth about the Padres, saying every fan's team is the best until the scoreboard tells them otherwise.

I mean, I'm not super convinced. It looked pretty AI-generated to me, but I find that a lot of these models struggle when somebody has a beard. So I don't know, maybe that's the problem. I thought this was really interesting as well. So you know NotebookLM, how you can give it a document and it will go and create an audio podcast where you can hear people talking back and forth about that document. HeyGen just released a new feature that does that but with video. Turn any doc, link, or idea into a two-host video show with studio scenes, multicam cuts, and B-roll ready in minutes. Now I already have a HeyGen account and I've already created a personal avatar in the past, so I'm going to set myself as the host. And then for the guest, I'm going to use like one of their built-in avatars. I think I'll use Hyper and and we'll do Piper Business Sofa. And then for the actual content, I'm gonna toss in the Alpha Evolve paper. I'm gonna set it to 720p just for the sake of, you know, making it fast. And then we'll go ahead and generate this. And it made a 2-minute and 40-second podcast about us discussing Alpha Evolve. But we won't watch the full 2:40. I just want to see how it came out. We're diving into a fascinating new white paper from Google DeepMind about Alpha Evolv. It's an evolutionary coding agent designed to solve complex scientific and algorithmic problems autonomously.

It's essentially taking the idea of natural selection and applying it to code, and it's already seeing some huge wins.

For example, it improved on a matrix multiplication algorithm. Okay, we get the idea. It's basically taking what NotebookLM did and making a video version of it. And I mean, I don't know, any of this like AI avatar stuff where it's trying to lip sync and make these people look real, I just have a hard time being a fan of it in general because it just means we're gonna see so much more AI slop and like low quality slapped together content just thrown on the web. And I don't know, I'm not a fan of where a lot of this stuff is headed. Like, I think it's impressive that we can do it, but not super excited that we actually are doing it, I guess. Speaking of AI slop, LinkedIn actually added a button called Seems Like AI Slop. So when you're reading a post on LinkedIn. And if it feels like it was something that was generated with AI and just slapped together really quick, well, you can click the Seems Like AI Slop button and give LinkedIn a heads up about that. I am worried about people weaponizing this and clicking the Seems Like AI Slop if they just don't like a post when it's maybe not even AI. But I mean, if people use it in the spirit that it was intended for, I think it's a good thing. The company Friend, who made the wearable that sort of listens all day and can kind of offer you feedback and advice throughout the day, well, they just released a new version that that actually speaks back instead of just like sending you messages on your phone with encouragement. It will actually talk to you. The old one cost $100, the new one $250. And quite honestly, I feel like their launch video is one of the most dystopian things I've ever seen. Like, I don't like where this is going at all.

I'm just struggling. Do you still think about her?

Yeah.

When was the first time you felt this way? A few years ago. Just feel like I should have an answer now. It's okay to be confused. I don't think so. It's not bad to be gay.

You get the idea. In other news, the Trump administration is banning new Chinese humanoid robots coming into the US, basically over concerns that they can spy on us, which is a bummer because all of the best humanoids are coming out of China. Right now. And finally, on the topic of robots, Google did just release a new model for robots, which is actually really impressive. It's called Gemini Robotics ER2, and it's the AI model that sort of powers the brain of the robot to get it to do complicated stuff. And some of their demos are actually pretty impressive. Like here's a video of it figuring out how to put grapes in a Ziploc bag without squishing the grapes and then actually squeezing and closing the Ziploc bag, which is a lot harder for a robot than you probably think. Here's a robot managing to unscrew a light bulb without actually like crushing the light bulb. Here it is tying up a trash bag, which is supposedly one of the most complicated things. I was on a little pre-brief call with Google where they were showing some of this stuff to us and just explaining how difficult of a task that is for a robot to do. So some really, really impressive stuff that we're getting to with robotics, and this is a topic I want to talk more about in future videos. So you'll probably hear a lot more about advancements in robotics and humanoid robots from me in videos because it's something that I'm like the most excited about right now. I'm less and less excited about like image and video generation models and AI avatars that can speak for you and, you know, make fake podcasts and stuff. I'm really, really interested in physical AI and how this stuff's going to be implemented in the real world. Like, that's where my brain really starts to light up. So expect a little bit more of that from me. But that's what I got for you today. A lot of stuff happened this week, as with pretty much every week in the AI world, and that's what I'm here for. I soak it all up. I drink rip from the fire hose and keep up with all the news and all the new tools on a daily basis because I know this stuff is moving so fast and most people can't keep up with it all. So I've made it my full-time job to do that and then turn around and make these videos for you on Fridays where I break down everything that I think is important to know that happened in the AI world. If stuff like that is interesting to you, maybe consider liking this video and subscribing to this channel. I'm like this close to getting to a million subscribers and you can really, really help me out. It's my goal to keep you looped in without creating a ton of overwhelm for you. I'm having a lot of fun doing it. And I really, really appreciate you tuning into these videos, hanging out with me, nerding out with me. Things are only accelerating, and I'm gonna be here to try to help everybody keep up with it and understand what matters and what doesn't. Again, that's what I got for you today. Really appreciate you. Hopefully I'll see you in the next one.

Bye!