Claude Opus 4.8 Full Breakdown & Testing (AI News You Can Use)

Claude was already the fan favorite of many people. And now they released another model, Opus 4.8. Today, I just want to talk about this big release along with some of the things that it came with, like dynamic workflows where it spawns hundreds of subagents and gets a big task done. And while that, it eats up most of your usage in your Claude account. We'll look at that, we'll look at the model, and we'll look at consumer preferences. And we'll touch on some other stories that appeared this week. But honestly, this is the big hitter this week. This is what we're going to focus on. Let's begin. So yeah, basically new model from Anthropic, Opus 4.8. You can see it here available in the web app. It's available in Claude Co-work and Claude Code. It's available in the API. This was a quick release since 4.7 that happened recently. And the reason for that was 4.7 was probably the model with the most mixed reviews where people were like, I'm not sure this is better than 4.6. If you weren't following that story, 4.6 was the model that made all of this agentic madness possible that really unlocked Claude code that made people buy Mac Minis, that made OpenClaw work between Opus 4.5 and 4.6. Those models were what made that possible, why this whole agentic revolution for consumers started moving. And 4.7 was better on benchmarks, but it was a bit too precise. It took everything literally. So what they did on 4.8 now is a bit of a course correct. They said, okay, it interprets ambiguity a bit better than 4.7, which was a big feature of 4.6. All right, I don't want to talk about all these model numbers. I know it can get confusing. And I don't want to spend a lot of time on benchmarks. But I do want to show them because, you know, they're above and beyond what 4.7 did. And they're also above and beyond what GPT-5.5 does in many instances. But then try to make sense of all of these numbers. Just because these are higher doesn't mean it's better. User preferences is usually what matters. People on the internet talk about the vibes. And so far, the vibes on 4.8 are immaculate. People are liking it. But here's the thing, these benchmarks the company showed, they always show the ones that are in their favor. Look, they selected 1, 2, 3, 4, 5, 6. But not even these benchmarks tell the full story. And there was a benchmark that got a lot of attention this week that I want to show you, DeepSWE. And basically what they did here is they tried to create a real-world software engineering benchmark where tasks are written from scratch. So the model has no chance of training on those exact tasks. There's a high diversity, there's shorter prompts, just like in the real world, people working with these models don't always write these long elaborate prompts. Often they kind of just wing it and let the model figure it out. And more. And unfortunately, this does not have Claude Opus 4.8 on here yet. This story got some attention earlier this week before the model came out. But as you can see, this ranking represents the real-world reality of these models way more accurately than what these companies put out. So for example, GPT-5.5, a lot of people claim that once you get used to how to prompt it, how to use it, it's actually superior in these agentic apps than Opus. And it wins out here. Now I suspect that 4.8 will land somewhere around here. But rather than focusing on the competition between these two that we've talked about for the past few minutes, I want to kind of highlight Gemini 3.5 Flash, which overall is not very well received by people doing more intermediate to advanced things with AI. Sure, they use it in their new Google search and in Gemini products. But evaluation like this seems more accurate than Google's representation, which is like, hey, we're at the top of this benchmark, and it's just the best model and the fastest model right now. Sure, they'll have the pro version of this coming and we'll have a look at this again. But I just wanted to show this because I felt like this felt more accurate than all these blog posts that we're getting where they kind of toot their own horn. Okay, now let's talk about what came out with this. Not just that the model is different, and we'll do some comparison prompts here in a second, but also they introduced dynamic workflows. This is a thing exclusively for Enterprise, Team, and Max plans where, well, it spawns like 100 sub-agents and those go out and do something very complicated. I've already heard the rumors of this. But what I want to try in this video is let this thing run and see how much of my weekly usage this eats up. Because that's the big question, right? Even though I'm on a $200 plan, my suspicion is that this is going to eat into a 2-digit percentage of my weekly usage. And by the way, we get offered selection here, which before we didn't, we just got adaptive thinking, right? But you get 5 levels now, you can go all the way to max, it's gonna eat up more usage. But this is really neat. This is something their competitors had. Now they have it too. Let's just crank this thing. Let's go Opus 4.8, effort maximum. And I do like trying this prompt. I'm not a big fan of a lot of these test prompts where, you know, we're creating 3D games. I do them, but I'm just not the biggest fan. Whereas this one, create a visually stunning design website for a studio that will impress web frontend developers. It's kind of very subjective and open-ended. And each one of these models has a different kind of taste profile on what it considers to be impressive. Impressive to a web developer. So really want to see what we get here in a second and also how long this is gonna take. Okay, here's the website. This took over 10 minutes to build. Let's have a look. We make interfaces that lean oblique. Ooh, I mean, come on. Have you— this is just me personally. Have you ever seen an element like this or anything like this with AI one-shotting it? This just makes me want to move over. This is gorgeous. And you know what I like? This didn't work nearly as well with 4.7. With 4.7, it took things literally. So you had to tell it what kind of elements you want in there. And look at that. This one got the creativity back a little bit, just like they claimed. I mean, it's just one prompt, but I'm impressed. I'm sure there's like other stuff hiding in here. Look at that email. This is great. I'm going to be using this model more just so it can surprise me like this. So we can kind of hop on over to a second prompt that I just run, which is the classic create an SVG of the Death Star in the sky above Los Angeles. Just a simple visual benchmark showing what it comes up with. Honestly, very comparable to some of the other models. And then I tried it against Opus 4.7 on the max setting, which is totally overkill for something like this. But I just want to see. This is 4.8 versus 4.7. They're just different shades of— okay, that's a Death Star above LA. Hey, and if you're enjoying this content, make sure to subscribe to the channel. I make videos like this every single week. And now let's look at what's next. And then next up, I actually just tested Claude Code's workflow feature where it spawns, as they say, hundreds of subagents that do different tasks. This feature is really meant for large jobs like refactoring a codebase or a big migration, like these massive tasks. Now, those tasks don't demo really well in a video like this, because you kind of need context on the project to even understand what's going on. So I did a very simple thing. One thing that I like to try a lot with these agentic coding applications, building a personal finance dashboard. I put this prompt right here into Cloud Code on my laptop, like a few hours ago. And I used the keyword workflow. As you can see, it highlights in blue. In my prompt, which activates this workflow function that spawns all of these subagents. And I also changed the model to Opus 4.8, the 1 million token model. So it has very large context, and I just let it run. Let me tell you, for about 18 minutes, it was just building out the project and the task, and it was making a plan. In total, this ran for almost 45 minutes and used up 300,000 tokens, which I was actually surprised that on my max plan, that only amounted to about 4% of my usage. So I guess it depends on the task. Refactoring a codebase and it's going to run for 4 hours, it's going to eat up more. But here's the thing, as I looked over it, I was really impressed with how thorough it was with just really making sure everything works. As you can see here in different sections, it planned out everything feature by feature, then it double-checked its own work, and then it QA tested everything, meaning it went through all the different interactions, uploaded mock data, and made a QA testing checklist for everything. So all of these things that previously you would kind of need to do manually and have plugins and have certain MCPs that can kind of QA test and do all of that. It just does it. It figures it all out. And let me tell you, I couldn't find the fault in this application. I tried everything. I re-uploaded data with new numbers. It got added perfectly. Everything populated. The design is right. The mobile version is optimized. That was a step that it did there in the middle too. That's really exciting because with a lot of this AI stuff, you know, it does things for you. But a lot of times it's like 80%, 90% of the way there. With this feature, sure, it'll run for an hour, but it's actually done. So I think this might be interesting for a lot of people watching this video. And I would actually make this recommendation if you just want to build something from scratch. I mean, heck, if you do have token limits or you're not using your Claude account on that day, you can kind of just let this workflow feature go. As you saw in my account, I don't use Claude Desktop regularly. I mostly use Open Claude for my own stuff or Claude Cowork for a lot of workflows that we teach. That's a nice sweet spot. But with this feature and me having the Max account for testing purposes anyway, I'll be using this. I'll let this just run. I'll open up an extra app and I'll let it build things even though it's not a massive, massive project that might require this. But you know, I'd rather have it 100% done and go overkill on this option rather than 90% done and then me going back and forth with it and fixing little bugs. I thought this was really great. And I don't know, over the next week, I guess I'll try it on some more specific and more complicated tasks. And yeah, overall, between Opus 4.8 being really solid and this new selection in Claude where you can select how much effort it will put in. And these new workflows in Cloud Code. I think this is an amazing release. Honestly, go try it out now. It's available in all Cloud plans. And yeah, that's pretty much what I got to say about Cloud Opus 4.8. And then finally, amongst all the other stories that were on my radar this week, honestly, only two of them that kind of stood out. One of them was a reaction to last week's Google releases with, you know, AI search here and new model here and new app and all of that. And what happened in response is that DuckDuckGo, the competing anonymous search engine, installs surged by over 30% last week. People saw Google AI Search, so they moved to DuckDuckGo. Thought that was sort of an interesting reaction. And then as sort of an add-on story to that, Google customized their Google search results, like the AI overview piece at the top. And they're starting to customize it to you. It kind of picks up on your preferred sources. So each person will get a different type of AI overview. Interesting. They definitely have a long way to go in developing that. And some people are just not loving the immediate product changes. And then there were all these other stories too that I kind of looked at, but honestly, nothing stood out. Maybe ChatGPT for PowerPoint I thought was really good. We're creating a standalone video on that. Spoiler alert, it's the best PowerPoint AI thing that you can use right now. So that's pretty much everything I want to talk about with you today. If any of these other headlines on screen caught your interest, the links will be in the description below. My name is Igor Pogany, and I hope You have a wonderful week.