Thoughts after a week of GLM 4.6 using opencode. Y'know, I really like this style of agent. I like the personality, and fluid creative process.
I like it better than Gemini, but not quite as much as I like GPT5-Codex.
Reminds me a little of Claude circa Sonnet 3.5. Good for web work, can visualize with modal bleed, recursive depth isn't bad, and I could probably improve the performance of it with more advanced proompting.
I'm also experiencing looping issues with it. Especially when it's time to work with Docker containers. Problem here is that I can't tell if it's the model, the agent, or the client, and I'm not curious enough to dig for why it keeps happening.
I think this one would be better for non-interactive code generation projects, where the outline is granular, and the instructions are explicit.
I just don't like it in a Claude Code style workflow.
The context summarization step is nuclear. Lossy isn't the right word for what I'm seeing.
Mocks are everywhere. Ask it to write a training loop, and it'll be fake. It's bluffing in all the same ways claude does when it hits too much reasoning depth.
That said, I think it's actually one of the better models for QA that I've ever run into, and it's a better creative writer than GPT or Claude. And now that I've paid for the annual plan, I have an entire year to play with it, and huge allotments of coding time.
I think the next experiment is to give it to my daughter, and see how she does coding up her text based adventure games with it. She just experienced Gemini Pro having a nervous breakdown for the first time, and she found it rattling.
We'll see how she does with this one.
She's learning, so her apps aren't going to be as complex as mine.
And I'm curious to see what she does with it.