Codex vs Opus --- The New Era of AI Coding Tools

The coding wars is heating up as the big two released their Frontier models just 27 minutes apart. Opus 4.6 and GPT 5.3 codeex. The fact that they are 26 minutes apart or 25 minutes apart is just insane. Are you telling me that OpenAI is probably a bit mad at anthropic and they're trying to to show they have they have something to release? May we'll talk about that. Might just be a massive coincidence. So I I I maybe won't read into it too much, but essentially Opus 4.6 improves on coding skills. I mean, if you don't know what Opus is, Anthropic have claude, which is their model, and they've got three levels. Opus is the big boy, you know, costs the most, most uh capable. Sonnet is sort of in between, and then Haiku is like the fast, snappy chat model that they have. Um, and yeah, we were on Opus 4.5. We've now increased just just a digit. It's interesting as well that chat GPT increased just a digit as well. So I don't know maybe they didn't know. Not too sure. It plans more carefully, sustains a genetic task for longer and can operate more reliably in larger code bases. Has better code review and debugging skills to catch its own mistakes. And in a first for our open uh class models, open 4.6 features a 1 million context window and that's in beta. So very very uh capable uh like smashing all the benchmarks as we come to know at this point. I mean you know I mean a gentic pretty good increase there. And if we look at the numbers we have Opus released before um Sonnet and before Haiku which is normally not the case. Normally we have Haiku. Uh we have Haiku Sonnet Opus because Opus is the the best and the beefiest one. Yeah, indeed. Um wanted to bring this as well. It introduces agent teams. Although I think this is more of a clawed code uh feature, but agent teams being able to spin up um you know asking to build a feature and having a specialist agent on your front end, specialist agent on your back end, specialist agent on your UX and then they can combine to uh and and communicate with each other to work on the task at hand, which is pretty pretty cool. Um, [snorts] I was going to show quickly how to how to do that because there is a bit of a there is a bit of a knack to it. That's actually interesting. I want to know how to do that because I started um to plan an app. I have a document for it. And I started tingling with the idea [snorts] of what if I task list that I have. What if I could name the tasks um and apply them and assign them to different agents like agent 1 2 3 4 5 and then have like a way of running like five different agents and each would go and pick their own tasks and do it. Yeah. [clears throat] So, are you telling me this is going to do this presumably automatically? I mean, you have to I mean, it needs to be it's not just going to do it automatically. I think you you need to allude in the prompt to say that you want multiple agents. Do you know what I mean? Like, so the first thing you do in your in your settings JSON, which you can you can get in your root the root of your um file uh user um where are we at here? here claude settings JSON you got to turn claw code experimental agent to true or or one and then when you spin up claw code nothing will change in your prompt that you say build this x feature make sure to use separate agents and whatever so it's not just going to pick that up um or at least it it'll be more likely that you want to point it to do that rather than it to you know try and do it automatically. I think it's better that you are more explicit with it. So, that's kind of a cool feature um of it. They also have made some here you go. This is relevant to you. They've also made some um cyber security improvements uh to both of the both of the models as well, which is really really cool. Um, and yeah, I think this is bigger and better, which is why I wanted to kind of combine the two uh the the release of GPT codeex and Opus in that we there's not much more to say. They're just they're just better. Do you know what I mean? Ultimately, the same thing. Um, you the the only thing to note is again this was released 25 minutes after OP OP OPEX Jesus Christ Opus. Um, but it it scored significantly higher on the uh the terminal bench. Other than that, I think we need to look at artificial analysis, which haven't made that haven't it's not as up to date as as comparing them all yet. Again, it's just been released, you know. So, they they've got not got a good comparison between Opus uh 5 oh my god. Jesus Christ, there's so many numbers right now. 4 Opus 4.6 six and codeex uh 5.3. No one has the comparison between the two yet because they just got released so close apart. I think what when was it yesterday they got released or last night even? Last night. Yeah. So, the way I understand it is just just use the the newest model. You know, if you're a developer, you'll just look at the numbers and just use the the highest number. Especially now with Opus being having the higher number, it just makes it much easier in a way to see what is the best. That's the best. Just use it. [gasps] And and it's such a good such a good point as well because we'll get into codeex and the new tool that there was released alongside all this. You know, I just prefer clawed models. Like I'm I'm interested to, you know, going back to shiny object syndrome. I'm interested to see what all these latest models are doing, but I think by now most people have just sort of found their the model that they prefer, right? Um yeah, if unless you're encountering major issues or unless you're in a you know, it's crazy that this time maybe but maybe not this time last year anymore, but like just over a year ago, we had we were handwriting all of our code. Now we're we're just really being pedantic about minor things that an agent is just writing so much of our code. So anyone not really interested in discovering like the latest and greatest and and splitting hairs on all these different models, you can't really go wrong with all these things. I'm not about to tell you that Codeex is better than um Opus uh or Claude even. So yeah, in in an attempt to to not make people sort of too overwhelmed, they're both great models, [laughter] you know, um and it's good to know that there is improvement. So pick your battle. If you're already paying for GPT, chat GPT plus, keep using codeex. It's great. If you're losing chord, use cord uh cord code, then use cord, you know. Um yeah, so that's it. I mean that's that's the news. Newer models, new some newer things that again the there's a steer you can steer Oh, what are you doing? There's a you can steer the models um in GPT codeex the actual app which we'll go over. You can actually steer it like midway through. It's constantly checking back with you and checking to see if there's any additional um information added. You can kind of do it with clawed code. I have done it um you know during a little break or whatever when I've cued a prompt. It's kind of taken that into consideration. I've not needed anything any more um more responsive than that. But we'll see in the codeex app how responsive that is. Maybe you can show us a little um interrupt a prompt or something. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web.