Transcription
I'm running my OpenClaw cluster of AI agents 24/7, 365, entirely for free. And in this video, I'm going to show you exactly how you can do the same with a Frontier model. Not some kind of crappy Llama local model, but the latest and greatest Frontier model from the big providers.
So, if you haven't heard of OpenClaw, it's kind of been the viral sensation of the AI world. It's an AI agent framework. It's open source. You can install it on any computer, but I have mine running on a Mac Mini here over in the corner. Mine's name is Janet. And Janet runs on a model, right? It has tools, it has memory, and it has a model. That's what an agent is. And a model is something like Claude Opus 4.6 or ChatGPT 5.3.
And the problem with OpenClaw is when it first came out, you could connect it to your Claude plan. So, if you're on the Pro or the Max plan, you could connect it to that and use tokens, kind of the currency of AI, with that built-in plan that you had. Well, then people started getting in trouble. They would get banned and their accounts would get cancelled. And if my Claude plan got cancelled, it would be a really big deal because I rely on it for everything. So we had to rely on using tokens. And tokens by themselves, and you pay kind of per use, are pretty expensive, especially for the latest and greatest Frontier models like Opus 4.6 or Sonnet 4.5 now.
So a lot of people started playing around with cheaper models and open-source models from China, and those are all fine and great, but there's a better way. And shortly after OpenClaw itself and Peter Steinberger were acquired by OpenAI, OpenAI came out and said, "Hey, you can use your ChatGPT plan as the daily driver model with your OpenClaw." And that's exactly what I've done.
And for you to do the same thing to get the latest and greatest model from ChatGPT and OpenAI to work as the driver model in your OpenClaw cluster, you just need to run these three commands. The first, `openclaw onboard --off-choice openai codex`. And then `openclaw models set openai codex gpt-5.3`. Codex, it's the absolute latest and greatest model, super powerful, incredibly good at coding, and really just anything. And then the last one, `openclaw model status --plain`. Just to give it a kind of plain descriptor here.
And so what this is doing is when you install OpenClaw, you need to select a model. And a lot of times you'll choose Opus or Gemini or whatever. And here we're choosing OpenAI Codex GPT-5.3. It's the absolute latest and greatest model. And truth be told, I use Codex and GPT-5.3 a lot when I'm doing coding activities. I use both Codex and I use Claude Code for local activities where I want to build product or create email sequences or marketing copy or landing pages. I'm using Codex more and more for software development and using Claude Code and Opus 4.6 for marketing and ops activities. That's kind of the current state of things and that'll probably change in the future.
So if you want to stay up to date with the latest and greatest best practices and AI for business people and makers, make sure you subscribe to the channel here. But what this is doing is it's giving us access on our Mac Mini, or you can run OpenClaw in a VPS like a Digital Ocean droplet or EC2 instance. It's all fine. Has some security concerns when you're running in the cloud like that, but giving it the absolute best model you can does a few things for me.
One is it makes the results I get way better. Using a model like GPT-5.3 Codex just means that there's less issues, there's less errors, there's less me having to repeat myself or it going off in the wrong direction. It's more secure. So models like Opus 4.6 or GPT-5.3 Codex are less susceptible to prompt injection and kind of malicious actors potentially invading your OpenClaw cluster.
And the last is better models use less tokens. You might have noticed if you use Opus 4, Opus 4.5, and Opus 4.6, each time the cost per token goes up, but the total amount of cost to get a thing done has been going down. It's my experience at least. Maybe 4.6 is a little more verbose. I find it really likes research and sometimes we don't need all that. But I think what we're seeing is the models are more powerful. So they have to call things less times and use less tokens to achieve a thing.
So for $20 on your ChatGPT plan, that probably everyone that would watch this video has, you can get the absolute latest and greatest model to run your OpenClaw cluster and you don't need to spend any more money. And I've used all of them. I've used Mixtral K2.5, Mixtral Max M2.5, both through OpenRouter. I find that's the best way to run these Chinese open-source models for data privacy reasons. I've used Opus 4.6, Sonnet 4, and they're all great. But with this, you just get the latest and greatest model. You can use it for everything for $20 a month. And I've not run into any kind of limitations with usage in the last kind of four or five days that I've been doing this.
And shout out to Andrew Warner for this great tweet thread here talking about like a tweet from Peter Steinberger. Hey, OpenAI has said this is publicly acceptable. Claude and Claude Code and the Anthropic team has not been as clear about this, and so we don't really know whether it's acceptable or not. You might get your Anthropic, your Claude Code plan cancelled if you use it to authenticate into OpenClaw. So I just wouldn't do it. And in this thread, Andrew Warner walks through the steps that I just laid out here, why they work, why they're great, and what you can do about your Anthropic plan if you choose to move a different direction.
But these are the only three commands you need to authenticate your ChatGPT 5.3 Codex model into OpenClaw. As always, if you're enjoying this, please like, subscribe, smash the bell. If you want to hear more about OpenClaw in the future, let me know in the comments below.