Open AI Unleashes Codex AI; Powerful New Vibe Coding Agent

So I really wasn't expecting OpenAI to drop a brand new research preview today, but here we are. They had a little bit of a live stream this morning and now they are introducing, or should I say reintroducing Codex, a cloud based software engineering agent that can work on many tasks in parallel. Meaning you can kick off more than one task at once and watch multiple work. And this agent is powered by Codex One, a brand new model built on OpenAI's O3, but obviously more to tuned into coding availability, chat GPT Pro team and enterprise users today with plus following soon TM, whatever that means. Don't worry guys, I definitely want to test it out today. I've been refreshing my chat GPT at the time of recording this. Right now I don't have access just yet, but I do have BRO resubscribed because I'm actually really excited about this. Back in the pre chat GPT days in the OpenAI playground you used to be able to have access to the code DaVinci models. Obviously they weren't great, I mean nowhere close to what we have today, but there actually also was a playground where it could generate JavaScript that you could interact with in real time. So OpenAI has been looking into AI powered coding interfaces for a long, long time now and this really is a build off of these older systems. So yeah, here is the new interface. Obviously this one's a little bit more advanced. You could see actually where it's looking in the repository for the code. In this screenshot example, this is, you know, the user's current project. Find a bug in the last five commits and fix it. If it can reliably fix bugs, that could be huge because it would be fully automated. This isn't just a new model from OpenAI. This is a new agentic framework for AI coding. Anyways, under that you can see different tasks that you might schedule to the model and then some older archived ones. Scan the entire repository and flag any variables, parameters or properties whose and then we can't read it. But you can see in this particular task it removed 27 commits and added 57 or I believe they are lines. And then this whole thing would be, you know, it's like a series of commits, right? I don't know, I'm not super good with the coding lingo, so forgive me, but you can see what it's adding and what it's subtracting. Very clear and well outlined. A huge difference with the interface. I mean I've tried to vibe code before and as someone who knows nothing about coding. The older way of doing it in like cursor or something like that is pretty confusing. If you are already a coder, it's easy. But for newbies, people who want to code with just natural language using AI, this is going to be the big ticket item. I think. At least until Google or another company collapse back Quick pit stop. You ever watch ChatGPT? Cough up 50 almost good video titles and still feel stuck? Same I started testing Spotter Studio, today's sponsor, because it's built only for YouTube nerds like us. On Spotter Studios home you can workshop any idea. So I tossed in a recent video title to see if I could upgrade it. I kept a few good ones and then tossed the rest and those specific good ones I was able to modify. And and it's all based off of your YouTube channel data. They even have a slider to make it more like your style. The IdeaBank auto tags every single spark and slaps a best, better or okay score on it. My wishy washy list of notion ideas. Yeah, that's gone when it's published. Time Outliers shows which AI topics in shorts are spiking right now at this moment. This is something I previously never cared about because it was far too much work and I simply didn't have the time. Spotter Studio has automated this. This can definitely boost CTR and it's a total freebie for me now. If this looks like a game changer for you as a content creator, they're running a pretty good deal here at 99 bucks for a year of Spotter Studio access. And of course, if you want to try Spotter Studio, the link is in the description below. Test out the seven day free trial, see how it does in some wild YouTube niches. If it spits out anything interesting, feel free to leave it in a comment below or send it to me on Discord. All right, with our AI toolbox fully restocked, back to your regularly scheduled content. The value of what we're looking at today goes pretty far beyond just a new model release. It's all about the interface and how it works under the hood. Since Codex is just a research preview, we're accessing it in chat GPT for now. We can assign it new coding tasks by typing a prompt, of course, and then clicking the code button. But if we would rather ask a question about the code base not, you know, make changes, we have to click the ask button. Each task, when you send it off, is processed fully independent in a separate isolated environment. Preloaded with your code base. So it's essentially duplicated and it's running it in like a dummy environment because if something goes wrong, we don't want to screw up the OG code base. Now in the environment, this agent Codex can read edit files as well as run commands, including test harnesses, linters and type checkers. And for those of us who are new to coding, I had ChatGPT whip something up. A test harness is like a mini obstacle course you build so your code can run laps while you watch for any face plants. It'll feed in some fake data, collect the results and let us know if anything trips or fails. We want to make sure that the code Codex one is generating is actually going to work in our project. Right? ChatGPT actually puts it better. Saves you from shipping code that blows up in the wild. Think crash test facility for software. Anyways, Linter is a grammar and style checker for code like the squiggly red lines in word that nags you about sloppy indentation, unused variables and other bad hygiene standard type of thing that you'd want in your ui. To be honest, this is like a precursor to the test harness. If something is screwed up and the Linter notices it, well, you're gonna want to fix it, obviously. Finally we've got the Type checker, which ChatGPT refers to the bouncer at the club. Every variable shows its id, for example an integer or a string before getting in. If you hand the bouncer a banana when the dress code says numbers only, you're out again. These are all kind of save fails to double CH your code. We want to make sure this stuff works. OpenAI says that task completion for code typically takes between 1 and 30 minutes. And I like that. That's a longer time frame than we're used to getting with AI. Hopefully it does a lot under the hood when it gets to a 30 minute autonomous agent Codex runtime. And of course you can monitor the AI agent's progress in real time. So once the agent completes a task, Codex will commit its changes in the environment. Codex will also provide verifiable evidence of its actions through CIT of terminal logs and test outputs, allowing you to trace each step taken. When a task actually is completed, you can review the results, request further revisions, open a GitHub pull request, or directly integrate the changes into your local environment. Apparently the Codex environment here is actually configurable to some degree to match your real development environment as closely as possible. Not sure how true that's going to be. So you Coders let me know in the comments. Another really interesting thing here is that codecs can actually be guided by a file you place in your code repository agents MD file so it'll automatically be looking for that. Basically like a readme for AI. You can inform Codex how to navigate your code base, which commands to run for specific testing if you already have them set up. And overall, how to best adhere to your project's standard practices. We want to treat it like a real human developer, give it plenty of context and information so it performs at its best. They did do some evaluations which I don't know, I kind of think they're pointless. Only getting comparisons to other OpenAI models here. But Codex 1 is a little bit better overall than SWE bench verified then oh three high. What's really great is that we actually get some extra information for the product that we're going to be receiving today in chat GPT. Codex 1 here was tested at a maximum context length of 192,000 tokens, which is pretty significant. Honestly. That'll be decent for coding. Not going to lie, especially if this scaffolding, this environment that it's built, the model to work in is slick and works well. It's also set to medium reasoning effort. So that's what we're going to see today in the product. I like that they actually tell us what's going on. Some more claims for the Codex 1 model under the hood. They wanted to train it to align outputs closely with human coding preferences. When compared to OpenAI03, Codex1 consistently produces cleaner patches ready for immediate human review and integration into standard workflows. Again, we don't really know because this is just coming from OpenAI. The community will certainly tell us if this is true or not once access starts being doled out a little bit more broadly. They do have a few comparisons here. We can see that the Codex answer is a little bit more tight, it's a little bit neat. The OpenAI031 seems to be a little bit more wasteful. And yeah, that's kind of what I'm picking up on so far here. The code is written in specified little batches to specifically address certain pain points where OpenAI03 consistently just generates a giant blob for you, which, I mean, if you're a good coder you'll be able to use it, but if you don't know anything about code, this will definitely be a little bit more confusing. We've also got some early use cases. These are basically just quick demo videos. I Think these will be very useful for giving us an idea of how to use codecs and how to actually get started with it. So I'm going to check some of them out. Take a peek at this one first.

My name is Calvin. I work on the Codex team as an engineer here at OpenAI. It's come together really quickly, but the end result that we're launching into the world is something that frankly is pretty magical in my opinion and has sort of changed the way that I write software. As an engineer, typically when engineers are running services, there's one person or a couple people on the team who act as the on call. So they're kind of the first line of defense as like a firefighter for the team. And actually I'm that this week, so it's particularly top of mind for me. At any moment I might get paged and have to abandon whatever I was working on and context switch to this new thing to figure out what's going on. There are a couple of things that Codex really helps with here. Oftentimes when you're on call, you'll get some sort of alert and you'll kind of look through the error logs and you'll see, oh, hey, there's an error at this one particular part of the code base. And sometimes it's something I'm familiar with, but sometimes it's not. When the minutes and seconds matter in terms of responding to an issue, the first thing I'll do is just fire off that stack, trace into Codex to understand where the issue is happening and see if it can fix it. Usually I'll like do that as a first line of defense and often the results that comes back are pretty good. I'm actually getting paged right now. This alert has been a little flaky though.

Wow, dude got paged in real time.

Another thing that I use Codex for is tuning our alerting. Right now. Since we're about to go into launch, we haven't fully tuned this to where it needs to be. Sometimes we get alerting and it's kind of a false positive where you used to kind of work as an individual engineer in like you'd have your task and you'd be writing code. Now it allows you to work almost more like you're your own team or like a manager of a team. I can just send it over to Codex now I can just consider it done.

So at OpenAI itself, this seems to be very, very much useful. But it is also OpenAI's own product. There's going to be costs for businesses associated when integrating this, I mean the cost of the product itself as well. And for individuals. There's also the fact that a lot of people that are Vibe coding with AI are used to cursor. They might not want to switch over to this. A lot of people really like Gemini 2.5 Pro for coding and if this is worse in any way, someone's going to point at it and say, hey, don't want to use this now. Because Gemini 2.5 Pro is kicking it for me. Which I understand as I've been mentioning, the big deal here really is that interface difference. It's not all about the model, it's about being that one stop vibe coding shop.

My name is nacho. I'm an iOS and iOS engineer at OpenAI. One way that I've been thinking about the interaction of human computer there's two battlenecks. Sometimes my brain is the bottleneck. Other times my hands are the bottleneck. Sometimes I'm working a hard problem and I'm just struggling to think about it, right? My brain is the bottleneck. I need to spend more time whiteboarding or thinking about the solution. But other times my hands are the bottleneck. Like, I know I have the code in my head. It's relatively straightforward. I just need to actually write it. I can use AI to work on things that maybe they're trivial, but I'm no longer limited by the speed of my hands. Codex actually helped me begin the integration of codex on the ChatGPT iOS app. And I wanted Codex to set up the scaffolding. For example, I asked it to create a separate Swift package. Codex set up the basics of the project. So by the time I got to my desk, I had that started for me and I could get started on the actual features. One thing I want to be able to see is archived Codex tasks. So I'm going to type that right now and I can just tap send and it's going to get the process started for me. I can actually follow along the progress with a live activity. As a developer, I can only work on one thing at a time, but if I know I need to do five things, I can get them started. And so by the time I'm done with this one task, I already have the starting point for the next one. It is another engineer in your team. I'm excited for developers to realize that we can do more and faster. This is already a massive change in productivity and how I work makes me.

Excited to try it for sure. I mean, he said that he wanted to start a brand new project. So he kind of explained it in natural language and it went ahead and got everything ready to start adding features. One of the largest pain points I noticed attempting to Vibe code as someone who isn't really a native coder at all was the fact that I had no idea how to get things started. I didn't know a lot of that root basic stuff, setting up certain folders, installing certain things that's going to make the project work cohesively. If this can handle all of that through natural language, I don't know. This could be the best Vibe coding interface. There definitely are some other ones that are good at automatically installing certain dependencies and taking care of things for you, but none of them have been able to do it without me having to open up another ChatGPT window somewhere and trying to like ask it questions about coding and all this other stuff. If you don't know the basics, getting into Vibe coding is definitely like, you gotta watch a YouTube tutorial at least. So I'll be interested to see if this removes that takes it a step farther. All right, we've got one more to take a look at.

This feels like one of those kind of transcendent moments. My name is Max Johnson. I'm a software Engineer here at OpenAI working on ChatGPT codecs. This, this is a special launch. I've never actually worked on a product that I felt as connected with as this one, because I use it every day. It's solving a very real problem that I have in my work. When I'm working day to day, I'm constantly seeing problems with the code, little paper cuts, things that don't meet best practices. Or maybe they're little bugs that maybe they matter, maybe they don't really matter, but may bite someone in the butt later on, you know, if we're not paying attention. And if I go down and try to solve all these problems myself in the moment, I would just never build anything new, Right? Nothing would get done with Codex. What's really cool is when I see these paper cuts, when I see these little problems, I don't have to break my attention. I can fire them off into Codex and come back to them in a, you know, 30 minutes or an hour when it's done. And slowly, over time, the quality of the code base goes up without any.

True AI agent behavior.

Let me show you an example. So recently I was working on a piece of our code, which is sort of a testing harness for our backend, and I noticed the code had a Few problems where I felt, you know, could be better, but I don't know if they really rise to the level of me interrupting my main work, starting a new branch, a new pr just to fix these things. So instead I'll just fire them off into Codex. I see that there's a retry strategy here, but it looks like part of the logic was deleted as part of a refactor, which is unfortunate but common. So I'm going to go ahead and go over to Codex and I'll just say in this file, it looks like there was a regression in the request JSON method retry implementation. I'll just say review the behavior from before the problem was introduced and then reintroduce that behavior and fix the bug. There's so much to build, and these little chores sometimes feel like they're standing in the way of the most important and impactful work that we could be doing. And so being able to come off to a model is super empowering. I've never felt more engaged or excited to be an engineer than I do right now.

Cool. This is a pretty great way to just kind of look at everything on the surface. We've got all the different tasks that ChatGPT is doing. You can click on them to get a deeper look, but it's all obviously occurring inside of a certain repository. And you can see the lead off here where it's branching off of as well. So you can kind of contain the AI to where you want. It definitely gives you some more controllability when you're working on specific issues. I'm going to be really intrigued to see if I could just like pick this up and start coding some things. Just start building with this. They're also making some updates to Codec Cli, which runs in your terminal and can actually operate in your machine. There's a new, smaller model of Codex, one that is built specifically for this, more similar to like a Mini or a Turbo model. Now, obviously the community is really hyped to get their hands on this and start playing around with it. One big criticism though already is that plus users do not have access right now. And that is really aggravating. There is a massive price jump of 10x from plus to pro just to try this new feature out. You have to fork over a ton of money. I mean, it works out for OpenAI. At the end of the day, I understand why they do it, because they can really. Plus, users are going to get access soon, but it's like you pay for plus to get early access to stuff and then you just don't Pro gets it instead. So kind of aggravating. But yeah, I did indeed pay the extra 180 bucks to get access to Pro just to try this out out because I am pretty hyped for it. I was originally an early tester of OpenAI Codex all the way back in the day. I want to see if I can just build games and build useful applications without actually having any real coding experience at all. You guys can make fun of me all you want for not being a coding expert, but the whole point of AI is to bridge gaps like this. No longer are you going to be limited by not reading a bunch of textbooks and spending years learning how to code to actually build something that breathes life into you. Right? We want to build things that change the world and shape it in new and interesting ways. And just because you're not a huge fan of sitting down reading textbooks and spending hours upon hours of your time coding, you can take a good idea and turn it into something at the very least. So here I am in Chat GPT. If I go to the model selector here, you know we've got all our typical models. I'm on the Pro plan. No access right now. Shucks. By the way, I am far from the only one who doesn't have access, but we should be seeing it sometime today. For the time being. I'm going to pause this recording, start editing a little bit, and then we'll come back and test it out. Hands on. All right guys, so I've been waiting to get access to OpenAI's new codecs research Preview. It appears I still don't have access, but here's how you can definitely check if you do have the Pro plan. As Nick Benson points out, he was able to get Codex access as soon as he connected his account to GitHub on the browser and it will appear as a button by Sora and Operator so it won't be in the standard model picking interface like I have demonstrated here. We actually do have proof of this from this user. Alex, you can see Codex appears over here in the corner. You can see I did set up my own GitHub repo for Codex testing specifically and I have connected my free GitHub account to my ChatGPT through the settings but still over here in the model picker interface I've got nothing. This video is already pretty long and I'm running out of time here, so check my Twitter follow the discord for updates in real time as I do get access and start to play around with it and expect a hands on deep dive video around Monday. For now, I want to get this out to the public as an info dump. So thanks for watching. I'll see you in the next one. Goodbye.