Over the past few days, I’ve picked up my old project: a Python library for a low-cost RFID reader I’ve been experimenting with. You can grab one for under €50 on AliExpress along with a hundred tags for just a few more euros. It’s a Chinese-made UHF reader, which means it can scan multiple tags at the same time.
Unfortunately, there’s no Linux driver available—only a C# driver for Windows plus a sample app from the manufacturer. So, I had to create my own cross-platform version in Python.
The library is still incomplete, but yesterday I faced the task I enjoy the least: writing unit tests. That’s when I decided to see how some of these much-hyped AI coding assistants would handle the job (you know, the ones that promise to replace “help” developers).
Here’s how they did:
- GitHub Copilot: A complete disaster—couldn’t produce a single test that actually ran.
- Claude.ai (Sonnet 4): Generated tons of tests, 100% coverage, looked great… but at least 12 of them failed. After more than an hour of back-and-forth, it managed to fix most of them.
- ChatGPT (version 5 thinking): Couldn’t fix Claude’s broken tests directly, but after a lot of conversation and going through them one by one, it gave hints that helped me fix the remaining ones.
- Chat.z.ai (GLM-4.5): The big surprise. At a price of €0, it delivered better solutions than ChatGPT-4 and nearly matched ChatGPT-5. Supposedly it’s the best Chinese LLM at the moment (ranking above Alibaba’s Qwen3, Moonshot AI’s K2, and Deepseek r1). I was impressed by the quality—it solved problems Copilot missed entirely, Claude tangled up, and ChatGPT only matched… and this was the free version. Impressive.
The takeaway:
- If you’re a junior developer using these tools, you might get yourself into trouble you can’t easily get out of. If you’re senior, they can sometimes save you time (Claude cranked out 53 tests in 2 minutes), but other times you’ll be wrestling with fixing their mistakes—or relying on your own knowledge to solve what they can’t.
- Right now, it’s hard to justify paying for Copilot, Claude, or ChatGPT while Chinese models like Chat.z keep shipping such powerful tools for free. And if you don’t want to send your prompts to China, you can still run them locally with ollama or LM Studio—provided you have an NVIDIA GPU with enough VRAM—since these are open models.
By the way, after trying VSCode, Windsurf, and Cursor, I’ve ditched that ecosystem entirely and I’m having a blast with Zed. It’s as fast and pleasant to use as the old Sublime Text but comes with tons of plugins and built-in AI assistants (next step: see if I can hook it up to Chat.z.ai).