You can replace Claude or Codex with a Chinese model/harness. But should you?
My first serious piece of work with GLM 5.2 and ZCode was neither a complete disaster nor a slam dunk. But what is Europe to do?
If you follow me on LinkedIn, you’ll have seen how bullish I am on Chinese open source or open weights LLMs. In particular I love Qwen3.6-27B running locally on my GPU fossil at home. All of my OpenClaw and Hermes agents currently use Qwen as their operational model.
You may also have seen the tons of hype about the massive frontier models coming out of China. The timing for GLM 5.2 from Z.ai was absolutely ideal as it landed smack bang in the middle of the Anthropic Fable drama.
Those of us in the EU who are deeply concerned about our over-reliance on closed US models, that could be cut off at any moment, started looking east during all of that drama. More on that at the end of this post.
Benchmarks Schmenchmarks
The story the benchmarks told was that GLM 5.2 wasn’t quite at the level of GPT 5.5 or Opus 4.8 but it was extremely powerful and probably good enough for many people’s use cases. It also costs a small fraction of the US models.
I tend to pay very little attention to those benchmarks. The only way to really find out how good a model and its harness is, is to use it in anger on a real project.
One of my favourite types of project to throw at these tools is to build an Admin Application for some service or set of APIs. The general shape of the projects is as follows:
Front-End in React
Back-End in Node.js
All deployed to AWS using CDK/CF
Running as Lambdas or ECS Fargate containers
Copious use of Cloudwatch
Using a range of DBs from DynamoDB to Postgres to even SQLite and DuckDB
Taking advantage of as much of AWS’s capabilities as possible
Talking to one or more APIs/Services of varying vintages, capabilities and documentation
As I’ve created many of these Apps in the past two years, I tend to point Codex or Claude at one of the existing ones and tell it to use that as a starting point. I add as much documentation as I can find for the new target system. And I lay out a pretty straightforward set of requirements. Most of the time both Claude and Codex do a superb job generating a usable v0.1 and we iterate from there.
Here’s One I Prepared Earlier
So for this project I took the same approach but using Z.ai’s latest ZCode Agentic Harness and GLM 5.2 LLM in Max mode. I even made it easier for them by pointing to a previous Admin App for the same system that Claude created, which I didn’t like. The fault for that older one lay entirely with me. Bad requirements lead to bad apps. But functionally it showed how all the APIs worked and how the overall object hierarchy hung together.
I kicked it off in YOLO mode where it didn’t ask for any permissions, and I let it loose on the problem. 1-2 hours later it said v0.1 was ready. Deployment failed miserably. Several intermittent hours later, the deployment finally worked. It seems GLM is not as familiar with CDK and AWS as the US models.
The code also failed miserably. I couldn’t login because it had created a Catch-22 in the code. Several more intermittent hours later and I had something that looked great but only kinda-sorta worked. Relentless Auth issues concerned me.
I threw a few more hours of tokenmaxxing at it and frustration levels rose. Every bug that was fixed uncovered more showstopper bugs. Eventually I’d had enough and exited ZCode.
Can you fly this plane, and land it?
I opened up Codex and gave it a prompt along these lines:
This application was built by another LLM.
It’s an Admin App for the XYZ APIs. I am very concerned about code quality and security.
It has the following limitations and bugs that I know of……
Let’s see if we can get it functional first and then do a full security review.
In less than an hour Codex had solved all the major bugs and fixed multiple UX problems.
I then got it to do a deep security review and it found many P0 and P1 bugs. Of course this was running against dev APIs and synthetic data on personal infra so the blast radius was minuscule. But it was disappointing.
It also took Codex less than an hour to solve all of those issues.
And the result is a genuinely useful Admin Application that I will be using regularly from now on.
Car Crash or Fender Bender?
As for ZCode and GLM 5.2? I’m making it sound like a disaster but it is far more capable than anything we were all using in Summer 2025. It’s just not at the level of Summer 2026. If they iterate rapidly on the model and the harness, it’s going to be absolutely useful for serious meaty work. My sense looking at the thinking output in ZCode is that the harness is more of a problem than the LLM. It just needs more time to mature.
If you are building apps that only ever run on your laptop, the combo is both powerful and extremely cheap. You probably don’t need Fable or GPT 5.6 Sol for that utility you bashed out yesterday.
China In Your Hand?
The wider question of relying on Open Source Chinese models is a tougher one to answer.
For personal projects, it’s a no brainer. Do it.
For Enterprises in the EU, when the model is hosted by a local provider and the harness is open source too, then I think it is reasonable to use them.
In fact, it would be prudent to get them proactively approved by internal compliance departments and added as fallback options for dealing with future Fable situations. Always have a Plan B.
We cannot bet the future of AI in large European organisations on the whims of another jurisdiction. Whilst China, as was rumoured yesterday, could cut off access to future frontier models, there is nothing they can do about already released Open Source ones.
As long as the EU itself is happy being the GDR of AI, then we’re going to be reliant on the charity of strangers.