OpenAI's New Coding AI is Insanely Good & More AI Use Cases

Welcome to this week's roundup of everything important that happened in AI. Coming at you from my grandma's guest room in Slovakia, Eastern Europe, visiting family for a few days, but that's not going to stop me from telling you all about GBD5 Codeex, which for many people was reason enough to switch away from Cloud Code. Interesting. Gamma 3. 0, which is an overall of the application that most people considered the best tool to create presentations within up until now. and so many more stories like agents getting access to credit cards and more. In this week's episode of AI news you can use, the show that takes all of the releases in this crazy space. We filter by what matters and then I get to show you live demos of how these things actually work in practice and how to use them. Let's begin. Okay, so let's have a look at

OpenAI's codeex CLI with the new GPT5 codeex model. Yes, the name is getting a bit more confusing again. I guess now they're just doing variations of GPT5. So I have this pre-installed in my computer. The setup literally took me less than 2 minutes. I linked it to my OpenAI account and now I can use the new GPT5 codeex model within their codec. If you're still confused by this, this is their version of clot code. And if you're not familiar with cloth code, you might want to check out the 70 80minute tutorial that I did about a week or two ago on this channel introducing you to that, showing what the possibilities are, showing you how to use it. It's essentially a command line agent and has become the go-to tool for agentic coding. And now with this new model, it's promising to be even better than before. And let me tell you, the vibes all over the internet and the sentiment towards this are extremely positive. A lot of people are jumping ship from claw code to try this out. But then a lot of it seems to be based on the frustration they've had over the past few weeks with clot code as it was delivering inconsistent results. Now that's sort of a whole different story, but as we're talking about it, I might just touch on that. Enthropic released an article about the clawed API models being inconsistent and giving their reasoning and everything. You can kind of check it out. But long story short, they're essentially like, "Hey, okay, we messed up. Sorry about that. " But I think quite frankly, a lot of the damage has been done and a lot of people have explored some of the alternatives. And now apparently Open I really went for the kill with this new model release, which we will test now. So what I'm going to do is my classic Pomodoro timer test just to get a feeling for this. I'll open up a new terminal in folder number one, claude pomodoro, and open up a second one with codeex. And then over here, I'll simply say codeex. Over here I say claude. Give it permissions. Proceed. Beautiful. Now we have the new codeex with the GPT5 codeex model active. And over here we have claude where we'll switch the model to opus to make this fair. Like so. I'll just give it this prompt. Build a simple pomodoro timer app with start, pause, and reset buttons. Show the countdown. 25 minutes by default. switch to a 5m minute break after each session and loop the cycle. Use clean, minimal UI and we'll hit enter on both of these. And in a sec we'll have two applications as created by these two state-of-the-art command line CLIs which we can judge. And while it does that I want to forward some comments that I've heard from some users. For some people codeex is unintuitive. They said they way prefer cloth code and I really wonder why that is. My experience with Codeex so far has been very minimal. I'll be very honest on this. What I did notice is that usually Codex went ahead and asked a bunch of questions before starting which I didn't get right now. It went straight into just doing. Matter of fact, what I see here is it's executing on the plan and while Clot Code asked one time for an extra permission, Codeex just did it all. And yeah, right now it seems like we have the application. So, let's see what the results are based on what it told us and then let's review it. Okay, it outlines the features. Okay, to try it out, open index. html in your browser. the timer will keep looping focus break sessions and then it makes suggestions for further features. I kind of like that cloud code as per usual gives you a list of all the features that it did and that you can find it in this HTML file. Let's say open the HTML. I mean heck, if you're using a command line tool, might as well take advantage of it. No need to double click anything. So here's the version from claude. This is a very cloud-esque look actually with the graded background. And then here is the version from codeex. Let's just have a quick look. Start. Okay, it's running. Pause. Reset. Okay, pretty straightforward. Works as expected. OpenAI version. Same thing. Works as expected. Okay, so how about we add a more complicated feature and see if that works. I'll just say implement multi-timer functionality and adjustable durations. That's nothing crazy, but that's quite a bit of code right there. Oh, there goes the claw version. Finished it first. I was predicted that I might want to open it. I like that. So, while Codex finishes, let's give this a shot. Yeah, let's just do deep work. 20 minutes. And I add a timer which I can start. Let's do another one with 10 minutes. And I add another one. And now these are running simultaneously. Yeah, seems like it can start this one and kind of pause this one. Aha. And now as I pause the second one, the 10-minute one stopped. Interesting. So, that's a bit of a bug right there. No, it seems like only one of them can work at a time. I mean, I guess that could be a feature, not a bug. Wait, but now none of them is working. Okay, it got close but not quite there. Obviously, we could easily fix that. But let's see if GPT managed to oneshot this. Okay, so we have the base timer. We can kind of just start this. As per usual, Claude was ambitious. Do you see how it did so much more with like adding names to it and that interface where you can like hide them away? kept it simple, which I think is an advantage in this scenario. Can just adjust this to 20 minutes started and yeah, this actually works. Look, the simplicity paid off. Ooh, I'm kind of seeing this now and liking it a lot. Yeah, this one just works right out of the gate and it didn't add any features that I didn't ask for. It's kind of nice to have all these buttons and stuff, but yeah, my experience with CL code is it's incredible, but often it's a bit over ambitious and you have to reel it back a little. So, there you go. In this quick little comparison, yeah, GPT5 Codeex amazing. And I hope this quick little test gave you a bit of a feeling for what these products can do. And if you still haven't tried them out, really seriously, go check out that clot code tutorial. It'll get you started with essentially both of these. It is in claw code, but most of the concepts there work in both of these. And this is just the most powerful tool in the AI space to learn these days. So, I hope this little comparison helped. Let's see what's next. Next up, we have

something I wish I had a bit more time with, but I guess we'll just do the same test we did in the Codeex CLI earlier on. Well, we just built the simple Pomodoro app. I just copied that over, but Replet released agent v3, kind of an overhaul of their no code building product. And even if you don't have a plan, you can kind of just try this thing out. Log in with your Google account, give it a prompt, and see what it comes up with. And there it is. Our first Pomodoro timer iteration. Let's just give it a quick spin. Start. That works. Okay, it added some features here. Ooh, settings and stuff. I didn't ask for that. So, this one is ambitious, too. Not a bad thing, per se. Ah, it's not even done yet, but it's like kind of works. Sound notifications, too. Oh, hold up. This is actually kind of cool. Look, it looked at the result and said, I can see there's a Pomodoro timer implementation, but it's more complex than what you requested. It kind of read my mind. But look, hold on. Yeah, it added statistics, settings, multiple brake types, and it's fixing itself. So, yeah, just like before, Replet is more of a autopilot tool that you can guide rather than something where you go step by step, which is really good in many scenarios. Look, the new version is as simple as I would expect it to and it's really easy to host and this interface just makes a lot of sense. Obviously, there's many competitors in this space, and in this channel, we've featured pretty much almost all the competitors, but honestly, this is a great option from what I'm seeing here. And I guess their agent free product is really looking to take a bigger bite out of the market. It just takes a while. I mean, I've been sitting here for 10 minutes and it's not even done yet. It seems to work already, but it just keeps working and working. Meanwhile, with both Codex CLI and clawed, I already had the extra features and I could have took like a 8 minute walk, but hey, speed is not the priority with these things most of the time, so that's fine. So, next up, we

have a brand new video model and this does something we haven't seen before, which is HDR. It's the Ray Free model by Lumi. And as it happens, this video is actually sponsored by Lumaai, which I absolutely love because we would have covered this on the show even if they didn't reach out to us. So, you might already know Lumi from previous episodes. This is their dream machine, which is their AI video generation platform. And now they have this brand new rayfree model, which excels in their visual quality. I would say it has this very artistic and tasteful style to it. And it's really good at HDR outputs, which means high dynamic range. In other words, that means the difference between the darkest point and the brightest point can be absolutely massive while everything being still visible. So, I have the product right here and I did what we always do on this show. We ran our standard AI video test prompts through Dream Machine with the new Rayfree model enabled down here. And right now, I'm running a second pass of our test prompts with the resolution being 720p and HDR enabled. This is going to give us the ability to really compare the difference between what a normal video output would be and these HDR outputs. And as per usual, we'll also contrast these with some of the competition, which is fierce in the AI video space. Okay, so let's have a look at some of these standard outputs and see how they did. Okay, so right off the bat, you can see that this model uses a lot of motion. This orbit around the sloth is quite the movement. I mean, the camera moves 90° in this case. The water looks really solid, which is something I always look for. And the consistency on the sloth is good, too. If you start looking really closely for some artifacts, you can find some in the leaves, but this is pretty damn good. And this shot of the sloth in HDR is just better. Look at all the rich details in the shadows here. You can see the entire sky and the highlights in the water, whereas in the standard version, when you look into the shadows, it's just black. That's pretty damn good. Let's have a look at the next one. Okay, so the water is actually really good. And I love the way it did the steam. This one actually looks better than some of the competition. Character consistency is good, too. And overall, you just have this feel that is quite artsy. I really want to hop on over to the HDR look here and see how that went. So, on this one, it's not visible immediately. I think we need a shot that has a bit more contrast. Maybe we'll look at the ghost shot. So, this is the SDR shot. And this one is particularly tricky because of the reflection, right? It needs to get it right in both spots. I love the general feel of this one. While it gets the eyes wrong at certain points, there's an atmosphere to this model that I really like. Okay, what about the HDR version of the ghost? Oh, this did way better. See how you have all the rich details in the shadows and still get the window and the ghost. And here the consistency between the mirror and the face is actually super good. Okay, one last one. What about this man with the dog? Pretty good character consistency. I'm actually surprised by that. And here the background looks perfect. Comparing to some other models, the humans on these look super good. And the HDR version actually does some lighting changes, which might not be what you expect, but you can really see the contrast differences between these shots and the SDR shots. Yeah, that delivers. One important note is that you can even take SDR footage and then within Dream Machine you can make those shots HDR enabling a new type of workflow where you can even take normal footage and just add the HDR effect. Interesting. So yeah, that right there is Ray Free. Go check it out from the first link in the description below. And thank you so much to Lumi for sponsoring this video and making these productions happen. All right, let's see what's

next. So now we got to talk about a new image editing model interface. Sure, it's an image generation model, but I don't think that's a special part here. competitors like Nano Banana just outperformed this on the quality. But what's really unique here in this release of Ravez image model and the interface that comes with it is the interface part. Now look at that. You get to generate various parts, various layers and then combine them. So you could do something like regenerate just the handback and then move the hand back around in the scene. And we went ahead and tested this to see how well this works in practice because this seems like the obvious next evolution of these applications. you really want is a AI generation and editing Photoshop like application, not one where you brush over something and say, "Hey, change this. " If you want to move the back, just grab the back and move it to where it is. I mean, that's the ideal consumer experience. So, yeah, as I mentioned, we tested this and we did a few things like replacing objects. So, you can see we first generate some lifestyle shots which look beautiful. I mean, this is a great image model, don't get me wrong. But then what we get is not just one image, but all these layers and these cutouts. Now, if you want to make this mug black, you would kind of just edit those elements right here. And then when you're done, it's going to regenerate. And there you go. The mug is black. And that's really kind of it. It's that simple. I should say that the same thing goes for moving objects in the scene and for changing text in the scene. It has never been simpler. I really, really like this interface. I mean, look at that. If you're doing professional work with this, if you're designing ads, posters, whatever, you just grab the damn product, move it where you want it to be, or just change the text to what you prefer. Great interface. In a

similar category, but maybe a bit less exciting, is the new model Cream 4. 0. It's another image generation and image editing model similar to Nano Banana. So what we did is run a bunch of comparisons to Nana Banana here. And yeah, here are some of the results. Adding this bowl of lemons, it's just different variations of good. No real opinion on this one. Changing the season in this shot, I think Cadream was actually more accurate. And in this one, where the model was asked to not change the aspect ratios, Cream changed it slightly while Nano Banana kept it consistent. And when we ran our image benchmarking pros, we just found that yeah, it's just another good model that does text. The stars look good. Honestly, a lot of the new models, they just look very similar and you can expect kind of the same things. It's really this editing ability that is special. One thing that I will point out is that when it comes to editing text, Cream performs even better than Nano Banana. Look at this one where we want to change the name from Horizon Labs to Parallax Labs with a double L and Cream got it right whereas Nana Banana forgot a letter. Something to consider. And I guess in this example, it changed the suit to a dark green suit, although the image is black and white. So, the prompt appearance here is even higher than Nana Banana. That would be my conclusion. Great model overall. If you're doing image editing with AI, try it out. Okay

now let's have a look at the one that I personally have been waiting for because when it comes to presentation creation, there seems to be consensus amongst all the people I talk to that the tool to do it with is this application called Gamma. They just built the AI PowerPoint from the ground up and it actually works really well. This is not affiliated. I always disclose when that is. It's just one of those tools that I really enjoy having at my disposal. And now they released a brand new release, GMA 3. 0, which overhauls the interface. And I myself am actually curious to see what happened here and what I get with the $8 plan I think that I'm paying for. So apparently there's really two big changes here. One of them is an agentic interface where you can talk to it and form your presentation that way. They sort of had this before. You could prompt over the whole presentation but I guess now it's going to be more interactive. And they're saying there's 10 times more visual variety which is amazing. I love that. Yes. Yes, please. And the second big release, which I'm not going to be covering here, is that they have this available through their API now, which can be really interesting. Having an API endpoint that can automatically generate slide decks for you. I mean, if you're in sales and do anything related to that, that just seems extremely powerful. You could take client calls and they just deliver custom slide decks after each call. This is a thing they even pointed out in their release tweet. But as someone who does a ton with meeting transcripts of all sorts, team meetings, client meetings, etc. This just makes a lot of sense. Okay, let's get into this and look at the new interface, Gamma Frequenc. Zero. What do they have going on here? Just going to create a new presentation. Generate it fully. Let's do it on cats with hats cuz why not? 10 cards. Generate outline. Okay, so this is still pretty standard. Looks very similar to what we had before. I guess the design variety will come in the presentation editing. Let's see. Cool. All this content works for me. We're kind of just testing. Yeah. And then themes. Okay. So, this is where some of the new ones come in. Let's see. Okay. Here on top there's a few new ones. Ooh, I kind of like this one. Or spectrum. This looks nice and modern. Uh, actually I'll settle on Mercury. Okay, good choices right there. I'll do line art images and generate the whole thing. Cats with fats presentation coming right up. Okay, I love the way it's created. I'm not sure that this was here before. I mean, I do use this semi-regularly. It might have been, but it feels different. And there we go. Cat Fats, a whimsical journey from story book to reality. All right, whatever. We're not going to judge the content of this. I just want to have a look at the design, which is actually really good. I mean, even down to the images, this did get better. I can't put my finger exactly on how, but like all of these are good and they align with the style. And there's a few new things. So like there's an agent here on top. Okay, I'm just going to say make it more crazy. That's always a fun one. And every request is free until October 1st. Okay, it was pretty cool. Let's see how that goes. Okay, so it updated the copy to be more crazy. Ooh, these variations. Okay, so basically took all the copy and just made it super crazy. Unleashed a hattacular Feline Frenzy. Okay, not sure I'm a fan of that copy change, but hey, that does work and it's what I asked for. Yeah, I mean overall this is a clear win. There's new styles, but the main thing that I noticed here is that just the design and the image generation, everything, it just feels more cohesive. I guess under the hood, they just started using more context while generating the images and they must have updated the pipeline that creates these presentations because to me, all these slides, they just feel a bit more coherent and like experienced designer made them versus before often it got you like 60% there, but you had a lot of work left. Whereas with what I saw here design-wise, I would say it's probably 90% of the way there. a very welcome upgrade. So already a great app that many people are using. So yeah, if you have any presentations, I can only recommend this. Let's see what's next.

And similarly to Gamma 11 Labs, another one of the big applications in the space, released a 3. 0 version of their studio application. Again, an amazing product that is getting even better. So let's just log in and see what changed here. So, here are all the generators as before, but they added this voice remixing, which is very simply explained a way to take a voice and create a variation of it. Not recreating completely from scratch, not uploading brand new training data to create a new voice, just remixing one you already have. I was actually going to mention that in the end, it just popped up here. So, now you know. But that's the studio release. The studio release is right here. And the point here is they just added a bunch of features which you might have found in other apps into here. And now you can find all of the different tools in here, too. So for example, this AI script generator is new. It's not something we have never seen before. It just wasn't available here. So if I say, write me a short poem, cats with hats, it will just generate that for me and then I can keep working all in one space. And as you can see, this seems to be a trend right now. Like all of these different applications, they're bringing in features that just make sense and bundling it into these like studio experiences where they want you to stay on their platform from beginning to end. Okay, let me switch this to the Unreal Tournament Voice, which is one of my favorites. Um, if you're not familiar, that was like a shooter game back in the day where they just had this legendary announcer with the most epic voice imaginable. And I guess if I select that, I could use tools like the voice changer or moving background audio right in here. And then all of these other functions are in here, too. I could be adding sound effects. So, how about I add a cat meow? That's a good one. And then I can import it here, and I can place it in the timeline. Okay, I see. Uh-huh. And then here, I can add a 1. 5 second break. Maybe shorten the sound effect. And now it should be that poem read in the Unreal voice with a meow between these two lines. Okay, let's see if it worked. — A feline friend, a dapper chap — with a jy hat upon his nap. That kind of worked. So, see, you can use this as a editing software, layering in things like music. And they even allow you to add video on top here. And if you add your own audio, they have these features like caption generation that you would want in a video creation tool. So yeah, they just brought together a lot of the things they had. Added a few new features and that's it right there. Kind of cool to bring a lot of these magical I think that's really the word AI tools together. I mean, back in the day when we were playing Unreal Tournament in school, I would have never thought that I'll be taking the announcer's voice, generating cat poems, and putting AI generated cat sound effects in between that and then doing that for a living. I mean, what a job this is. But hey, I guess this is how we learn together. All right, let's see

what's next. Okay, next up is something that excites me personally because it's really at the core of what we do here with this channel and with the show and that's looking at how people actually use AI. What are the use cases? How do they put it to work? Two new papers this week. One from OpenAI, how people use CHBT and the second one from Enthropic and their economic index. Particularly the Enthropic one is really fun to explore because they did this interactive interface. You can go by state. So you could check out, for example, California over here and what it breaks down to. Most people use it for computer and mathematical use cases. You could also group by category. Number one thing providing educational tutoring and academic assistance across multiple platforms. Nice. Software dev is in second place. And then you have things like IT support, edit documents. Yeah, this is just really great. What about Florida? What's the number one there? Software dev assistance. And then the documents. And then interesting ones like provide personal relationship advice and life guidance support. Almost 1. 5% of the use case is pretty significant. I just find this extremely informative and easy to navigate and open released something similar with a bit of a different approach like how many of the chatdy queries are related to work. The answer is the later the user signed up the less they use it for work. So work was what really drove a lot of the early adoption and now these days only about 30% of the usage is for work and more than twothirds of it are for personal use. Interesting that and much more in this document. Go check them out if this interests you. I know I found a lot of data points here that I'll be using for my own educational work. Whether that's creating courses in the community or helping companies succeed with this stuff because with papers like this, you can base the education on real world experiences rather than just feelings. And I love

that. Okay. And for this week segment, let's start by talking about this Albanian AI government minister. I mean, honestly, I don't know what to make of this. I just wanted to bring this up. He's supposed to handle public procurement and they say he's the first member who's not physically present but virtually created by artificial intelligence and the whole reason is that he's 100% incorruptible. I have so many questions here like how is he incorruptible? Can you just like prompt inject him and then who will be using him? Who's prompting him? Is he prompting himself? Who sets that up? I don't know. I tried to find like all these details on a story. I couldn't get them. I guess I'll just keep an eye on this. But yeah, this is happening. There's like AI ministers popping up

now. Then we have Google leading with something they call the agents payment protocol unifying all these different companies behind a standardized way for agents to pay. And I think this is actually really interesting because it's kind of an extension of MCPS which are an open standards to use tooling with agents. Then I think two weeks ago we covered the agents. md file introduced by OpenAI which is a standardized way to manage context with agents. And now Google is trying to do the same for payments essentially making it so that agents can use your credit card easily and safely. I mean look obviously that comes with a lot of problems right but honestly if you can unite MetaMask Mastercard all the Gentic apps Visa American Express and all the other big companies I mean look at this crap that's just crazy then this is what building a new internet that is actually powered by AI and agents looks like there's a public GitHub repo now with all the details on this nothing to do here really I mean if you want it you could like download it and try it but I guess soon we'll have standardized ways of doing everything with agents which is kind of great I'm always a big fan of interoperability and this open approach

Okay. And beyond that, we have this story out of OpenAI, which seems more of a legal reaction than anything else. Yet again, a chativity is going to be predicting your age based on your interactions. And then if you're below 18, it's going to introduce parental restrictions. This is going to be linked with the parental controls. And I think in theory, this makes a lot of sense. I do wonder how this is going to be implemented though in practice. And the one thing I personally really care about is for full-g grown adults to have access to models that are powerful and not lobtomized in any way, shape, or form. I mean, I think some of that is really necessary for the wider public. I get that. I just don't want to see this divide happen where the big corporations have all of the most capable AI and then common consumers get the 80% weaker version of that, which is a scenario that could easily pan out and parental control and safety are the perfect veils to kind of push these sort of limitations. And that's really everything for this week's episode of AI News You Can Use. To round out this episode, actually, let me pick up the camera. Let me give you a little bit of context on where this episode has been shot from, cuz I think this is pretty unique. Classic Eastern European architecture. All right, that's it for now.