Introducing GPT-5 Codex: Optimized Agentic Coding for Developers

--- type: transcript date: 2025-09-16 youtube_id: Gs0bMFcP9lw --- # Transcript: Introducing GPT-5 Codex: Optimized Agentic Coding for Developers OpenAI has just released GPD5 codecs which is a version of GPD5 that's been further optimized for agentic coding within codeex. Now one of the interesting things with codeex that you may have noticed over the past several weeks as well as months is that codeex is increasingly available in more and more places where developers work. So whether it's within an IDE, whether it's VS Code, Cursor, Windsurf across those different extensions within the web app that's adjacent to ChatgBT or within a CLI app similar to something like cloud code or being able to access this within something like a GitHub action where you can tag codecs to actually review things like PRs or actually try and address particular issues that you might have. So today we're releasing GPT5 codecs, a version of GPT5 that's been further optimized for agentic coding within codecs. This is specifically designed for the OpenAI series of products for codecs. This isn't available for the API or anything like that. GP5 Codeex has been trained to focus on real world engineering tasks and is equally proficient at quick interactive sessions independently powering through long complex tasks. its code review capabilities can catch critical bugs before they ship and is available everywhere you use codecs. Now I think one of the key aspects of this announcement are just the number of places where you can access codec. So whether it's from the CLI IDE extension and one of the really neat features that I will show you at the end of the video is actually how you can access this and continually have the same context across different environments. to say you start something within the web app, continue it on within your IDE and back and forth and share that context thread almost as if you were using chat GBT on your phone and then you're sitting at the desktop and you can just go and access that thread. Another thing that they touch on within the blog post is they basically touch on a number of the different products that are named codecs. The one thing that I think OpenAI is really going for is trying to create a little bit more of uniformity around some of their product offerings. Say for GPD5 while there are a number of different models under the hood and it just routes the query depending on the complexity. A similar thing here with codecs is while there are a number of different tools with the same name they ultimately serve the same process of being the place to go when you need Gentic coding whether it's for review or creating net new projects or implementing features basically across the stack and across the different environments. There's a unified term that they're going for with Codeex for basically anything coding related. It falls within this vertical. Effectively, if you pay for the $20 a month account, you'll be able to access this from the web browser all the way through to the CLI or within something like VS Code. Now, to touch on some of the benchmarks. So, GPD5 Codeex High does rank a 74.5% when compared to GPD5 high at 72.8%. Now in terms of code refactoring, this sees a significant bump and they do touch on this and this ties back to how the model was trained. They describe this as being trained on, like I mentioned, real world software engineering tasks all the way from building projects from scratch through to adding features, tests, debugging, performing large scale refactors or conducting code reviews. basically the whole software development life cycle all of these actual practical implementations of what a software developer does day-to-day that was really the focus in terms of the training of the model now in addition to this the model is more steerable so they do describe this as being better at adhering to the agent MD which is similar to something like cursor rules or the claude MD or agent MD there's a number of different standards that are out there and basically what those are a way to tap into the system prompt where if you have additional instructions of what you want the model to do or not to do or if you have different conventions or all of that, you can include those within the instructions within markdown. Now, in addition to that, they describe this as something that produces higher quality code. You can just tell it what you need without having to write long instructions on style or code cleanliness. It's inferred that the model is going to be better at actually writing cleaner code. Next up, in terms of some of the architecture on how the model works, Codex adapts how much time to think depending on the complexity of the task, it's very similar to the default mode within chat GBT where it will route to the respective model under the hood as well as specify how much time to think through each respective task. Now, the other thing with this is the model does combine two essential skills they described. For coding, there is pairing with developers in interactive sessions and also persistent and independent execution on longer tasks. And this was a criticism on codeex when I originally used it is it didn't seem to actually go off and be able to agentically code for a very long time. But what's different with this is they described that during testing they've seen GPD5 codecs work independently for more than 7 hours at a time on large complex tasks, iterating on implementation, fixing test failures, and ultimately delivering a successful implementation. Next up, a number series of charts that was pretty interesting is they did actually even optimize the model for the comments that are being injected within code. And that is something that a lot of different models do bias towards depending on what model you're using, they're very verbose with the different comments that they include. And sometimes some of the comments that are included aren't exactly high impact. There's been a lot of memes about sometimes if you are using an agentic coding tool that it might comment something that is blatantly obvious about what it's actually doing. Maybe you're naming a variable or doing a very rudimentary piece of logic. Obviously, you don't want to have that extra bulk and cluttering up your codebase with things that aren't exactly high impact. So, it is nice to see that they're actually iterating on the smaller things as well for the model. Now, in addition, they do describe that GP2 codec is a reliable partner on front-end task and they also mentioned that they made some updates to codecs to make it a better pair programmer. They've revamped, like I mentioned, some of those different ways that you can actually access codecs, whether it's within the CLI or within the Codeex IDE extension. Here is an example of codecs within the CLI. So, if you've used something like CLA code before, this will feel quite familiar. You have the slash commands. You also have the plan that it will break down. You can see it run through all of the different tasks. It can execute the commands to install things. It can GP and search for different files within your system. Very similar to something like cloud code. But what's really interesting with the story of codecs is just the ability of now being able to access this basically wherever you want. So if you're a VS Code person, you can access it there. If you're curious about the CLI, you can access it there as well. And what I really like about this actually is being able to have codecs on my phone is sometimes I have an idea for how I can potentially fix something or improve something. Maybe I notice a mobile bug looking at my website or something like that. I can quickly go ahead and send in and ask the agent to create a PR to fix something a particular issue that I might see where I might not actually be available to use my IDE or a CLI. So being able to have it from the CLI to the different VS Code versions all the way through to having it within the web app as well as having the continuity where you can start in one interface and pick it up in another. It just makes it that much more flexible to use. And as you see here within codecs, the nice thing with this, one of the favorite features that I have with this is you can actually specify to have parallel requests for one thing that you're asking for. For one thing, you can actually spawn off four different cloud instances going and trying out different variations of what you might be asking for. It just makes it that much more effective where you can review four different PRs quickly and you can see which agent actually took the more or less proper direction of what you were asking for. Now, in addition to the CLI being able to access this within something like cursor, VS Code, or the web app, you can also access this within GitHub. If you do want to have another set of eyes, whether it's on a PR or on an issue or whatever you actually want to have within GitHub, you can go ahead and tag codecs with the instructions of what you want it to do, and it will go ahead and accomplish that based on the context of what'sever within that repo. Now, in terms of pricing and availability, so Codex is included in ChatGpt Plus Pro, business edu, as well as enterprise plans. You're going to be able to access this today. All of the different features, and you will be able to access it within all of the different interfaces that I showed you. And one thing that I do want to call out here is they do mention that you will be able to access this from an API key soon, but it will be specifically for the codec. One of the interesting things with this model release is this was specifically designed for their products and even when they're releasing their API, I think this is going to be an interesting direction to see if other companies, whether it's something like Anthropic, actually follow suit because a lot of the coding applications that are out there, whether it's web app builders or some of the popular tools like cursor or a lot of the agentic coding tools that are out there do rely on obviously products, whether it's from OpenAI or anthropic. It will be interesting to see if some of these model companies do begin to reserve the right to some of their frontier models for their own particular use cases, whether it's within coding or other applications. But let me know your thoughts within the comments below. And otherwise, kudos to the team at OpenAI for the release. And if you found this video useful, please comment, share, and subscribe. Otherwise, until the next