Original video in Chinese.
Key Takeaway
- Codex is currently the strongest general-purpose agent: OpenAI’s Codex firmly holds the top spot in the general-purpose agent field, far ahead of OpenClaw and Claude Code. It is mature, ready out of the box, safe, and reliable, making it the best choice for most people.
- Four core advantages: 1. The desktop GUI is user-friendly (history + chat + preview in three columns), completely lowering the barrier to entry; 2. Computer Use capabilities are mature (controllable mouse and keyboard, operating apps and browsers), with both a built-in browser and a Chrome plugin solution; 3. Dual-layer security mechanisms (user permissions + Codex’s active judgment on high-risk actions) make it much safer than OpenClaw; 4. It supports networking across multiple devices (SSH chaining multiple computers) + the $20 Plus allowance is very generous.
- Why I recommend it: Codex combines “ease of use, security, and productivity” better than anything else, making it the most mature general-purpose agent product right now. I strongly recommend that ordinary people use it directly, without having to mess with Claude Code or OpenClaw anymore.
I guess OpenAI itself never expected it could turn things around with Codex.
Right now, Codex firmly sits at number one in the general-purpose agent field. That was the judgment I made earlier in the Knowledge Planet.
If you’re using Xiaolongxia or Hermès, you don’t need to keep messing around. Codex is the mature product—ready out of the box, safe, and reliable.
If you were previously frustrated by Claude Code’s barrier to entry, you can relax now. Codex doesn’t have nearly as many tricks up its sleeve, and it is much more forgiving when it comes to network access and payments.
In my view, OpenAI got four things right in this round. It was also because of these four things that I recommended it several times in the Knowledge Planet and told everyone to definitely give it a try.
First, they launched a desktop version.
For most people, CLI is still too unfamiliar and too intimidating. They may not even know what a terminal is, let alone all the commands.
By contrast, the Codex desktop version feels much friendlier. The left side is the conversation history, the middle is the current chat, and the right side is the preview. Whether you’re selecting permissions or switching models, the controls are all familiar.
Second, they launched Computer Use.
The foundational capability of a general-purpose agent has to be Computer Use. Think about it: if it can’t even operate a computer, what kind of general-purpose agent is it?
Codex can now see the screen, click the mouse, type text, operate menus and windows, and complete tasks through a graphical interface.
For example, I asked Codex to open Apple Music and play “Now Now” by Maokun Poet. It would open the app on its own, search for the song title, and then click play.
My experience using it during this period is that Codex’s computer operation ability is relatively mature now, and it can work through either a graphical interface or the command line. Especially when it comes to operating browsers, people were just discussing this in the Knowledge Planet a couple of days ago.
The reason was that Codex now has two browsers it can operate: one built-in browser and one Chrome browser. When it operates the built-in browser, both speed and accuracy are very high.
That’s because the built-in browser has a shorter path and is specifically designed for scenarios like viewing the UI. And when Codex needs to access a real browser environment, it operates Chrome through a plugin.
OpenAI’s arrangement of two browsers is actually quite clever. If everything had to be done in Chrome, the experience definitely wouldn’t be very good.
OK, having the basic Computer Use capability isn’t enough; a general-purpose agent also has to talk about safety. That was also the reason I never used OpenClaw before. In terms of permissions, Codex has two layers:
One layer is the permissions for the system and apps that the user grants.
The other layer is Codex’s own judgment. When it encounters a high-risk operation, it will actively stop and ask.
After all, it is operating the user’s computer, so having these two layers of permission settings is much more reassuring. Just by looking at this one point, you can tell this is what a mature product looks like.
Third, personal AI device networking.
Codex is already connected to ChatGPT, and all you need to do is scan a QR code. That means when you’re out and about, you can use ChatGPT on your phone to give Codex instructions and operate the computer at home.
However, Codex can control more than just one computer. Through SSH, you can also connect more devices.
This move is really smart. Having multiple devices is already very common. For example, I’m still using a Mac mini, a MacBook Pro, and a PC. If I can use Codex to call on all of these devices, turn idle hardware into productivity, and build an AI device network, that would really give personal productivity a huge boost.
Fourth, the quota is ample.
I’m currently subscribed to the $20 Plus tier. It’s completely enough for daily use. And I often find that just when I’m about to run out, the quota gets reset again.
I think OpenAI is this generous for two reasons:
First, they really are behind right now. They have no choice but to be generous, or they’ll truly fall behind.
Second, agentic AI really burns through tokens. Just look at how GitHub Copilot subscriptions have had to change the rules. In that situation, if you’re still being stingy, then no matter how good your product is, people won’t be able to use it.
The four points I just mentioned—launching the desktop version lowers the barrier to entry; ample quota lowers the cost of use; launching Computer Use establishes the product; and chaining multiple devices lays the groundwork for the next step, a networked personal intelligent device setup.
Based on these four points, I’m very optimistic about Codex. That’s why I said in the Knowledge Planet that OpenAI will sooner or later need to solve the positioning issue between ChatGPT and Codex. Although Codex can’t match ChatGPT in fame or user count, its ceiling is much higher, and it should stand on equal footing with the big brother.
OK, that’s it for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!