Original video in Chinese.
Key Takeaway
- How I’m dealing with Claude getting dumber: Claude Code’s performance has declined (the official team admitted it was a bug), so I downgraded from the MAX plan to the Pro subscription and switched to OpenAI Codex (GPT-5-High). To my surprise, it solved problems that Claude couldn’t handle, and I’m now running both in parallel.
- Model comparison: Codex is cautious, thorough, and strong at following instructions (good for architecture and backend); Claude makes reckless unrelated changes; Codex is about twice as slow but powerful, while Claude is better at the tool layer. They complement each other.
- Principle and outlook: I don’t pick sides. I use whichever is the most advanced and strongest. I’m looking forward to Gemini CLI, and to building products with three tools in parallel: Claude, Codex, and Gemini.
Recently, Claude Code has gotten dumber, and that’s actually brought me an unexpected gain.
To deal with it, I started trying OpenAI’s Codex. After using it, I found the quality unexpectedly good. Especially on some problems that Claude Code had failed to solve over and over again, Codex managed to solve them in one shot.
If you’re using AI coding tools, I strongly recommend trying Codex. Remember to switch the model to the highest setting, GPT-5-High. You’ll be pleasantly surprised.
Anthropic officially published an article the day before yesterday acknowledging that there was a problem with reduced model output quality. They found and fixed two issues, but they did not admit to any intentional downgrading.
How should I put it? Justice is in people’s hearts. Anyway, whether this was a bug or intentional downgrading to control costs, I think we all need to take some action.
For me, right now it’s Claude Code and Codex running in parallel.
I downgraded Claude Code from the previous $100 MAX tier to the $20 Pro tier. I used MAX because I was aiming for the Opus model. Before, to make full use of Opus, I chose the Plan Mode approach using Opus. In other words, I’d first let Opus diagnose and plan, and then let Sonnet execute.
Now that GPT-5-High on the Codex side is stronger, there’s no reason to pay extra for that.
To be honest, although I’m not optimistic about OpenAI, GPT-5-High really did help them claw back a point in coding.
After using it for these past few days, I’ve started feeling comfortable handing it more work. For example, I had it help me optimize the user login authentication and data retrieval flow for two products—this was an area that used to have bugs all the time. I’d wanted to do it for a long time, but I was worried Claude Code might not handle it and that I’d need to invest a huge amount of energy in debugging and analysis. Now it’s all been handled.
Thinking it over, there are two things that left the biggest impression on me about GPT-5-High.
First, its working style is more cautious, and its thinking is more detailed and thorough. It gives me the feeling of being diligent and not impulsive. It takes time to think deeply instead of quickly giving a preliminary response.
By contrast, the model over there has more of a “young hotshot” vibe—it’s especially reckless, willing to take on anything, the kind that just charges ahead. Anyway, if it messes up, it can just admit fault right away, right?
Second, GPT-5-High is very strong at following instructions. It faithfully does exactly what you tell it to do, no more and no less.
With Claude, I’ve run into this several times: clearly it was a backend bug, but it insisted on changing my UI. Honestly, at the time I was extremely, extremely annoyed.
So now, when it comes to architecture and backend issues, I hand them over to GPT-5-High to execute. The reason I still keep Claude Code around is:
First, Codex is still pretty slow in execution. It feels about twice as slow as Claude Code. Maybe that’s the price you pay. The model thinks more, so it consumes more tokens and more time. That makes sense, right?
Second, while the model is stronger, in the coding assistant layer, Codex still isn’t as good as Claude Code, and there are still many areas that need polishing.
Third, OpenAI hasn’t exactly been shy about doing this dumbing-down thing. If I remember correctly, the first time I really learned about the concept of “dumbing down” was on ChatGPT.
So for me, it’s not a matter of replacing one with the other—it’s a matter of using them all. I just open two windows, one on the left and one on the right, and let them both help me work.
Now the three giants—Google, Anthropic, and OpenAI—have all released CLIs. It won’t be long before Gemini 3 arrives. I’m really looking forward to its improvements in programming. By then, maybe I’ll be running three at once: Claude Code, Codex, and Gemini CLI side by side on the screen, all helping me build products.
My principles are very simple, just two points:
Use the most advanced AI productivity tools.
Whichever is stronger, I use that one.
I don’t pick sides. That’s completely meaningless, a pure waste of time, and stupid.
OK, that’s it for this episode. If you want to learn about AI, become a super individual, and find like-minded people, come join our newtype community. See you next time!