We Tested OpenAI's GPT 5.6 for a Month

**Greg Isenberg** (0:00)
GPT 5.6 is here, and it's a big deal. My friend Dan Shipper has been testing it for weeks, and on this episode, he shows you how you can use Codex with GPT 5.6 to run your personal life and run your business. His setup is really, really cool. And by the end of this episode, you're gonna learn where 5.6 works, where it doesn't work. You're also gonna learn how to use Codex in a really, really, really cool, interesting way so that you can make money, be more productive, and just automate a lot of that boring stuff. I loved having Dan on this episode. He's a fountain of knowledge and he shows everything. So enjoy the episode and I'll see you at the end.
Really excited about this podcast, Dan Shipper. I begged him to come on, he came on.
Dan, by the end of the episode, what are we gonna learn?

**Dan Shipper** (0:57)
We're gonna learn how to use Codex to run your entire life, and how to use it to build a compelling software business.

**Greg Isenberg** (1:07)
Mm, you have my attention.

**Dan Shipper** (1:12)
Should we get started?

**Greg Isenberg** (1:14)
Let's do it.

**Dan Shipper** (1:15)
Okay, so if you haven't used Codex, it is the OpenAI answer to Cloud Code, Cloud Desktop, Cloud, you know, Cloud Cowork. If you've used it before, you may have used Codex CLI, which I do not recommend. It is like, Codex CLI is trash. The Codex Desktop app is what the Cloud Desktop app would have been if you had just, they had just like thrown it out, and they were like, what is the ideal way to build this? Because right now, the Cloud Desktop app, it's like, you've got chat, you've got code, you've got co-work, and I'm always like, which one do I use? And Codex, well, now they have two tabs, but really, it's a much simpler implementation, much cleaner, and much more powerful, because it was built after OpenAI got to see, OK, this new paradigm of using an agent on your computer to do your knowledge work, it had already started to happen because of Cloud Code and Cloud Co-work, and I feel like OpenAI just fast-forwarded through all the messy stuff, and they were like, this is the ideal implementation.
And I feel like Anthropic has a little bit of the mandate of heaven right now, and OpenAI is quite underrated, and I think Codex is the reason why. So Codex with 5.6, which I find to be the most usable, most powerful, fastest model for, specifically for knowledge work and for coding, but like specifically for knowledge work, it's not fable level, but fable is like, it's like a tactical nuke, you know? It's like, it's so powerful, it's like actually illegal. And 5.6 is, it's more like, I don't want to use a weapons analogy, so I'm gonna move away from the weapons analogy, because it's not a weapon. It's more like having a Porsche. It's like great to drive around town. You can do anything in it, you know, it goes fast. It handles really well. It's really good for collaborating.
I think of Fable as being something that you have to, you have to have skill, like a lot of skill to use it well. 5.6, it just like works for everything you wanted to do. So anyway, that's all of like big way of saying, 5.6 and Codex is where I spend all my time. It is my operating system for work. I do everything from my emails to writing to, I'm actually starting to train models now, which I think is a new frontier.

**Greg Isenberg** (3:46)
I don't even know what that means. I want to get into that later in the episode.

**Dan Shipper** (3:49)
Let's get into it. Let's get into it. I would say like just a real, like a small preview is training your own model is the next step after making your own skill. If you've made a skill for something and it's not quite working, train like fine tuning a model is the next step.
Until recently, it was completely out of reach and not at all practical. But 5.6 and Fable to some extent, if you can get around the guard rails, makes it easy enough to make a machine learning pipeline and run it. That's everything from grabbing the data to making more synthetic data, to actually running experiments. It makes it easy enough to do that. I think it's now possible for non-machine learning engineers to start doing it.

**Greg Isenberg** (4:38)
All right. Well, I'm intrigued. I'd love to see what you got. You want to start sharing the screen?