The unexpected death of Codex transcript

It's not a particularly well-kept secret that I really, really loved the Codeex app when it dropped. And I want to be clear about this. I'm referring to the Codex app. I know it could be confusing because Codex was the first OpenAI model for doing code all the way back in like 2022, 2023, if I recall. Codex was also the CLI that they put out for working with agents back in 2025. Then Codex became a special version of the model that you could use with the CLI.

Then it became the app. Then it became the website and became all these other things. and the codec's name got kind of diluted, but I still loved the app. I loved it so much that I built my own open- source clone just in case it started to have regressions or other issues and so I could use it with other harnesses and models. Today, OpenI announced an overhaul of their app called Chat GPT work and with this an all new chat GPT experience.

I actually have it open right here, but you might notice something. It says chat GPT but the icon is codecs. That's because they stuffed all of codeex into chat GPT. The codeex app is now chatgpt. When I clicked the little update button in the bottom earlier because I was in codeex and I wanted the latest version, I ended up installing the new chat GPT and I'm not very happy about it. There are a couple things in this overhaul that are actually quite impressive.

But there's also an absolute disaster of marketing to be discussed here, too. And I'll be frank, I'm disappointed. I knew they were going to do this change for a while now because I had friends at OpenAI reach out to me concerned asking for me to help them jump in front of this rebrand because so many people were so concerned about it and it seems like they weren't successful. I have feelings about this and they're not very good.

I'm too frustrated to do a good transition joke. So, let's just roll the sponsor and then I'll dive into all of my frustrations. You've probably heard me talk about today's sponsor before. It's Blacksmith. But listen, okay, I really need you guys to check this out because it has fundamentally changed how we think about CI for my entire company. The numbers speak for themselves. We're doing almost a thousand runs a day on Blacksmith and the vast, vast majority of them are under 2 minutes long, even though the equivalent run on GitHub would have been significantly longer, sometimes as long as 10 minutes.

It almost sounds too good to be true. It's 60% cheaper and two to 10 times faster, but it just is. You sign up your GitHub or, you swap the runs online in your GitHub runner, and you're good. But it's so much more than that. If it was just the faster runners, it would be the easiest recommendation in the world. But once you see how much better things like their observability are, where you can actually dig in and figure out what's going on across your services, it's night and day difference.

I've had to use GitHub actions for a few specific things like, you know, publishing through the trusted OIDC flows through npm, and I forgot how much worse all of the tooling was. and the moment I went back to Blacksmith because I had a bug in some runner instantaneously was so much easier to figure things out. They also have an awesome monitoring flow where you can pick specific projects and workflows to monitor for specific triggers.

For example, if there's a thing that you don't want your agents to ever do and you have a throw in your lint rules, this can set up a system to ping you when that happens and keep track of where and when those things are happening. Can you guys keep a secret? I don't know if I should talk about test boxes yet, but they're also really cool. uses their existing infra for near instant spin up of Docker images to run your agents to do real work.

It's so cool. If this ad was longer than your CI, fix it today at soy.link/blacksmith. Good to have you back. First, I'm going to do my best to try and explain why they made this change and the cool things that were added. And then I'm going to rip them a new one because I think this is really stupid. And then it opens Safari for me to sign back in with. Even better. And now I have to scan a QR code in order to do the pass key off because it won't use a browser.

Thanks, Anthropic. I should not have done that because they removed work. That was really dumb of me. We're going have to do a little bit of magic in the edit quickly because everything sucks now. I'm sorry. Anyways, to show what was on my screen approximately 45 seconds ago, there were three tabs in the Claude app. You had the standard claude app, the home, which was for chat and whatnot. You had work, which was a thing we'll talk about in a second.

And you had code, which was a bad interface for cloud code, which is the cloud code we all know and love, but in a guey inside of the cloud app. That middle option that vanished was called work. And it was their attempt to take the things that make cloud code magical for non-dev tasks and make those accessible for non-devs. Sorry, it was called co-work, not work. My bad. And co-work's been moved to now be part of this section here where you have chat and you have co-work and separately you have code.

There was an interim window where they had three options though. They had home which was normal chat. They had co-work and they had code. And the concept of co-work is really cool because it turns out having access to your computer, the ability to run commands, computer use, and all these other things is valuable outside of traditional dev work. I know this because the non-devs on my team have been using it heavily for all sorts of different stuff.

With Co-work, you can use the files on your computer. You can use the browser that you have open. You can update your PowerPoint slides. You can do all sorts of worklike tasks that aren't writing software. The reason they built this is the power of Claude code as a harness, not as a tool for writing code, but as a tool for using your computer to get real work done. And it seems like OpenAI wanted to do something similar, which is why they have rebranded Codeex to be a feature in chat GPT.

Instead of the chat GPT app being for chat and CEX being for, well, Codeex, ChatGpt is now a UI for Codeex as well as a UI for work, which is a separate thing that doesn't appear to do anything. I just clicked and it shows the exact same stuff on the side here. But it should hypothetically speaking allow for non-devs to get some of the benefits that we see from chat GPT. You might be a bit confused cuz I mentioned before that this is chat GPT.

So where's the chat? We have work. We have codecs. Well, if you go down a bit here, you see chat is an option because chat is no longer the thing you use the chat GPT app for. Chat is now a popup in the chat GPT app. Needless to say, I think that's really stupid. And I think the people who use the chat GBT app for chat GBT are going to be more pissed than the people who use the codeex app for code. And now I want to talk about the cool things that have been introduced.

They have all of these fancy plugins for PDF management, spreadsheets, presentations, sites, which is a cool new thing that's pretty competitive with Lakebed Myloud, so we'll talk about in the future. And all these fancy plugins for Gmail, Slack, Notion, Google Drive, and lots of other things. Seems decent so far. You also have the ability to choose which folders or projects you're working in. And when you create a new project, you're no longer just saying this folder is what we're working with.

You are picking an existing folder in your computer. And you can even select multiple to be used in a given project. It no longer maps in my head the way that I'm used to with, you know, everything else. We make a new project and we click here to edit it. Why can't I pick files for it? Is this cuz I'm in work mode? It is. Great. So, if I switch over here and I add a project. Next. New. Oh, did they change this? That's annoying.

The last time I tested this, you could have multiple folders in a project, but I don't think you can do that anymore. Weird. A lot of the UX has been changing kind of rapidly and not in ways that I love. But the point of all this is that it's easier for non-technical people to use codecs to give it access to things on your computer and do things. I want to talk more about the good things. So I'll do my best here. Obviously the sites beta is included.

I briefly mentioned that earlier. It is pretty cool. The plugins between chatbt and codex have been unified. So there's just one layer for all of that. The computer use stuff had a massive improvement, including things like multitab and enterprise off for browsers, as well as much faster computer use. The way that Andrew describes the difference between the codeex mode and the work mode is the interface. Codex is built to show you code, diffs, and PRs, while work keeps those details abstracted away.

You pick one or you can move freely between them. When you want to quickly explore an idea or get an instant search result, chat is now in the app, too. With the instant and pro models, search widgets, and seamless handoffs into work or codeex tasks. The internet browser is much better at helping chat GBT work across the web with support for authenticated sites, multiple tabs, and file downloads. Your tabs will stick around, too.

It's also worth noting that the official chat GBT browser called Atlas was sunset as part of all of this with the expectation that you'll use the browser in the chat GBT app instead which uh to the five people who really really liked the stuff going on with Atlas I am sorry probably not the best replacement as I have mentioned many times about 5.6 Six. It is way better at computer use. So when you combine it with the new computer use capabilities of the app, it's really, really cool. 56 has better taste at front end, yada yada.

It can visualize data and things well, which is cool. Codex side got some updates including editing files directly in app, seeing and reviewing PRs without having to leave the app, accessing pro mode in chats, and then handing off the context to codeex. This is actually pretty cool. I do a lot of like talking about an idea with pro and then have to copy paste the context. Having a built-in handoff is actually kind of cool.

There's the allnew ultra mode for your most demanding work. Not really a feature with this uh codec staff. And the most important piece here, if you've grown deeply attached to the codeex app icon, you can keep it. I am one of the people that did not like this change and ended up going in, hopping into appearance, scrolling all the way to the bottom where I saw this tiny little doc icon where you can choose between chat GPT icon and the codeex icon.

And this is the only proof of Codex's existence remaining other than the little call out here. They do plan on doing more with this Chad GBT codeex merge, but they want to do it thoughtfully, not just smash two things together with a toggle and call it a day. Sure, there's also a list of nice small improvements. Much better performance. I've noticed this. The codec app has been overheating my computer a lot lately and right now I don't feel almost any heat coming from it at all because the new chatbt codeex implementation seems to be more effective.

My honest guess is that this is what they've been using internally for a while and the updates to the codeex app are just keeping it safe and porting things over without actually testing it. So it's not super surprising to me that the codeex app has had problems lately because I don't think they're using it. And the reason that Codex was good is because so much of Opening II was using it. Now they've moved over to this unified app.

I would expect it to be pretty solid and could continue improving over time. They also cleaned up settings. They made the model picker and slider better. I do really like this sliding between the options once you have the model picked. It's cute. It's not as bad as before. Faster, more reliable mobile connections. This is huge. The mobile stuff was so garbage. I basically gave up on it and just bullied Julius into building it for us for T3 Code, which we'll talk about in a bit.

New docs already has been better. I found that when I Google search for things like the CLI install for codecs, I can actually find it. Programmatic tool calling with code mode. Useful. Nice to see that they finally added that. And better video rendering on SSH hosts for when you're connecting to codeex over SSH on another machine. You can get video back through the SSH as a connection method now, which is cool. But yeah, I'm I'm sticking with T3 code.

Okay, I think I've done my best to cover the positive so far, which means it's time for me to crash the out a bit. Agents have gotten really, really good at building websites. They still struggle with one important piece, though, the DNS. It's crazy how obnoxious DNS can be, even just for deploying one small site. Once you're trying to do more complex things like set up subdomains for users that they can configure themselves, let users bring their own domains over, programmatic registration, and more, you start to run into a lot of problems, especially when you want your agents to drive these things.

DNS simples existed for a while with the specific goal of making DNS simpler. And I already loved using them for every service I built that required DNS that could be configured dynamically. If I wanted users to be able to register domains or sign up for things, DNS simple made it trivial to do. Their SDKs are still the best I've ever seen for managing domains, but they went a step further. The CLI is unbelievably cool.

It's crazy just how far ahead they are from every other domain management service I've used. Most of them have dashboards that take 30 seconds to 2 minutes to load. DNS simple has a CLI that responds almost immediately with whatever request you have or more importantly, whatever request your agents have. If you had to think about DNS zones more than once in the past 6 months, I implore you check these guys out. They really know how to do it right.

And if you use my link, you'll get $10 of credit when you sign up. Check them out now at soyv.link/dnsimple. As I mentioned at the start of the video, Codex had its problems in branding. Codex went from a special model to a CLI to new versions and treatments of a model to a desktop app. And the result is that to this day, when somebody says, "I'm using codecs," I don't know what they're referring to. It could be the CLI.

It could be GBT53 codecs. Maybe they found a way to host that old codeex model back in the original Copilot days. I don't know. Maybe it's the app. Hard to know. For comparison here, when I say I'm using cloud code, it's pretty clear what I mean. If I told you guys I was using cloud code, what am I referring to? Am I referring to a model, an app, a CLI? If I tell you I'm using cloud code, what do you think I'm using?

See how almost every single reply is the CLI? It's because that's what people think of when they hear cla code. No one thinks about a special cloud code model. Very few think about the app. It's clear. It seems like even anthropic doesn't think about the desktop app too much. Claude code refers to the CLI. Codeex refers to a bunch of different things and none of them are properly aligned and the result is confusion and brand failures.

But they started to fix this. The first step in the fix was killing the codeex models. Previously they were training specific versions of the models to be better at code. They stopped doing that because they folded all those behaviors into the traditional like main models. That's why we never got a normal 53. We only had 53 codecs and then they folded those learnings into 54 and we'll probably never see a codeex model again as a result.

That was a huge step forward. The next big thing they did was really lean into the app. It's been clear for a bit now that the codec is not their focus. It recently got some improvements, but a lot of those improvements were from me bullying the team into fixing things, threatening to file some slot PRs if they didn't, but there were a ton of just like broken UI UX things in it. Like when you open it, if the wrong model selected and you type /model too fast, it would fail because it was still trying to set up and connect to the MCP servers.

It was just like obviously annoying and bad UX in ways that were trivial to fix. We're talking like 20 minutes of work max, maybe like 50 to 100 lines of code and they just weren't bothering because the whole company had moved over to the app. And while that has problems and while I wish the CLI, which is the only remaining open- source surface, got a little better, got a little more focus, honestly, I understand why it hasn't been the focus.

And I'm just thankful that they've kept the app server in as good of a state as they have. If you're not familiar with what I mean there, the Codex CLI hosts a server that lets you connect to it through apps like you know the official Codex, sorry, the official chat GPT app in Codex mode, but also things like T3 code. So those things helped a bunch. The biggest thing that helped with Codex's brand though was that the app got good.

It was decent at launch, but it continued to improve. It kept on adding awesome things that were unique to Codeex. things like the computer use capabilities that are just not at the same level in any other tool right now. And all of that combined with how generous the subscriptions were and how good the models had gotten at coding resulted in a huge resurgence of the Codex brand despite their attempts to shove pets in our faces.

And all of this led to Codex growing massively. OpenAI couldn't stop bragging about it. They were seeing 5x or more growth every month consistently. And it had gotten massive. They were buying Super Bowl commercials just for Codeex. But here's where things start to spiral. As soon as Codex became default behavior for the model, when they folded the Codex models in, the need to fold more in became apparent. The reason that they folded Codex's training in to the main models is because the things Codex did were useful outside of code.

Being able to do really longunning agentic work is valuable. It lets you complete tasks that otherwise couldn't be completed using LLM. I would argue that when 54 dropped and confirmed they were no longer doing codeex models, the inevitable end of Codeex was kind of painted on the wall. And as Codex kept getting better and better and chat GBT kind of I don't want to say faltered or anything, but it wasn't improving at the same rate.

It was hard to not stare at these two things where one is trashb the thing that makes your business most of its money that is massive and super popular and well used by almost a billion people and you have this much smaller thing codeex that is way more capable that's improving way faster that everyone loves. It's hard to not stare at these two things and be like, "Oh yeah, codeex that should be chat GPT." Because then they'll like it even more then they'll keep using and our growth will be focused in on where we are going.

And obviously all of the benefits of codeex like its ability to do things on your computer were very valuable as a harness helper, whatever you want to call it. The harsh reality is that Codeex was too good for Codeex. Since the team did so well making codeex awesome to the point where I know a lot of people who made the move entirely away from claude code and anthropic models, myself included for a while up until Fable.

Those people made the shift for a reason and it was because codeex had improved such an absurd amount. Obviously they wanted to expose all of that capability to chat GPT. My honest guess is also a bit of a team shuffling thing here where they had the goal of trying to get these incredible engineers building codecs, all of the awesome people they had hired from everything from marketing to Devril to to general comms and management of folks like myself.

Obviously, the engineers building the app as well as the insane people who are doing all the Mac computer use hacks, ex-Apple employees doing a lot of that, by the way. It's super crazy. I get why they wanted to fold that into the cash cow of chat GPT. But that on the topic of the company structure, I do think they've had some issues with the breakdown between research, apps, chat, GPT, and dev stuff like codecs. And I think that this announcement from Fiji is a pretty damning issue.

To the credit of OpenAI, it seems like she was having a lot of severe health issues that are the main reason she has been taking leaves and is now moving to an advisory role. But this is the individual whose role was previously the CEO of applications at OpenAI. She was brought in in an attempt to focus in on the product side as its own business and really make great apps like the chat GBT mobile app, the chat GBT site and then as well codecs.

And it seems like they've struggled a lot with doing this correctly. And she has personally struggled quite a bit too which is why she has decided to leave. It does not appear that she's leaving on bad terms. Sam Alman's even said that he's really sad about it, but is very grateful for all that she has done at OpenAI and even grateful for her friendship and who she is as a person, wishing her a speedy recovery. Again, health seems to be the driving issue.

But when the driver has health issues, you now have a driver issue. And since she is the one in charge of driving OpenAI's application development, this isn't great. That's also the mistake because part of why Codex was doing well wasn't just that it was so good. A big part of why Codex is doing well is because it wasn't chat GBT. Codex users didn't want Chat GBT. If they did, they would be using Chat GBT. They wanted Codex.

A large part of that is the functionality that Codex has, but a lot of it is that it's its own surface built by and for devs, focused entirely on our use case and our needs as developers. I made a post earlier today about my feelings here. Codex rebranded to chat GPT. Claude is spamming me with Excel plugins. XAI is bragging about how good Grock is at PowerPoint. I knew the AI labs focused 100% of their effort on software devs thing, but it's sad having it all end at once.

This is kind of how it feels for me. The dev stuff that all the labs have been focusing on is still very valuable, but it's time for them to try and branch out beyond developers, trying to expose the awesome functionality of codeex and of cloud code to people who aren't just writing code. But I think they made a huge miscalculation by doing this. Not only were people using codeex because it wasn't chat GPT, they're also using it because it wasn't Claude code.

And as such, they started building an allegiance to it. Everything from people showing up at events wearing codeex merch to people getting really hyped to talk about it. I even know people who have been trying to make codeex specific meetups happen. Codex is a brand that went way beyond OpenAI in ChatGpt. It was a brand that felt built for developers and it was resonating with them a lot. It resonated with me so much that I ended up making my own clone of the Codeex app.

I try to not show my stuff constantly, but you know what? It's fully open source and free and doesn't make us any money. I'm going to talk about T3 Code a little bit. The genesis moment that led to me building T3 code was actually during the early testing of the Codex desktop app because I liked the Codex app so much that the random performance issues that appeared a few days into testing frustrated me deeply. So deeply that I wanted to make sure I could fix it myself if it broke.

And when they told me they did not intend to make the desktop app open source despite the CLI being open source, but that the CLI was how the Codeex desktop app worked through all of the app server stuff. I realized it might be time for us to make our own open source alternative so that if we do have these types of problems, we'll be good. We can fall back to an open-source project that can be modified by us, can be pinned to an old version, can be fixed by us.

Whatever you need to do to make sure T3 Code behaves how you want it to behave, you can do now. To be frank, as much as I loved the Codeex app, I didn't trust OpenAI to not screw it up in some way. Whether it's performance issues, whether the models just stop getting better and I want to use it with different models, whether they make some UI change that I hate or I guess kill it entirely. I wanted to make sure I had a solution that I could trust and that you guys could trust too because it's open source so you can do whatever the hell you need to it.

And that's why we have this steal our code legally button at the bottom because I want you guys to be able to use T3 code for whatever the hell you want and a lot of people have. There are forks of thousands of users and as many as a fourth of our users are running a fork or a patched version of some form and I love that. I want to support that. I want people to be able to change their editor experience however they want to.

And I'm more thankful than ever that we did this because I don't want to use this version of codecs. I just don't I don't like having all of these things I don't use all the way at the top here, especially on my smaller screen. I don't like that. The combination of that and pin chats means that projects is so far down the UI. Now, I don't want sites as a special button built into my UI here. I don't want chat as a popup window that happens when I'm just trying to write some godamn code.

I don't like this direction at all. And what I like even less is the erosion of the Codeex brand because it was starting to do really well. Codeex had grown to the point where it was an actual competitor to claude code and now it's a lot harder to talk about that and this is why I think they really fumbled here. Having codeex as like a little call out here is nice and all but it's not the same. It doesn't get people excited to talk about how they're using codeex instead of quad code.

Rero put this really well in chat. Instead of hey man you got to try codeex. It's now hey man you got to try the codeex mode in the new chat GPT desktop app. Yep. Even people like Jensen Hong, the CEO of Nvidia, sent emails to Sam Alman about how excited they were to use Codeex to accelerate and do work never possible before. Codeex, that's over. To put it simply, this is one of those changes that just hurts motivation.

Both the motivation of the people using Codeex to share it with others, but also the motivation of the team building it. I know a lot of people that joined OpenAI to work on Codeex, the coding experience. And now codeex is just a toggle inside of the chat GPT app. It's just a bad decision company organizationwise. This might make sense. And if you're the type of person who works at OpenAI and is thinking about these things constantly, the layers here don't matter as much.

But this is a really common mistake for people who work on one thing full-time. They lose their ability to take a step back and see how the world perceives it. And in this case, the world is perceiving this as very confusing. A bunch of people opened up their chat GPT app today to have it turned into a code editor with a chat button and a popup shoved into it. A lot of developers installed an update for codeex just to have the app replaced with chat GPT.

And now when they tab around it says and shows the chat GBT icon instead. And maybe if they go find this hidden menu, they can make it at least look like the codeex icon even though it says chatgbt under it. It's just dumb. This is just not the right path. It's a fundamental failure to understand how people think about these things and talk about these things. But at least they haven't taken the app server yet, which means other tools can still use codecs without leaving behind developers.

Because my fear now is that the things we need as devs are going to be depprioritized as chat GPT has new features they want to shove into this top left section. I've never been so thankful to have done a sloppy open-source project as I have with T3 Code. I still can't believe how far it's come. And I'm so thankful for people like Julius making it what it is today. It was a hailmary bet that was kind of hedging on the possibility of Codex getting worse.

I will be real. I was not hedging on the possibility of Codex becoming a toggle, but here we are. I don't think I have anything else to say on this one. I am just sad to have lost one of the brands and apps that I genuinely really liked. I let OpenAI know ahead of time what I thought about this change. They didn't seem like they wanted to hold it back and I guess I wish they did. I have nothing else to say about this.

It's a sad way to kill a brand that was growing and was well loved within the developer community and I wouldn't be surprised if they started to see sentiment shift as a result of this seemingly simple change. I got nothing else to say on this one. I'm going to go play around with the new model because at least that's really good. Until next time, he's nerds.