Day 2 using GLM 4.7: This model is an astonishingly good code generator. Base personality is weak, and it doesn't respond to creative tangents without pushing back. I'm not happy with the alignment on this model at all. It's getting in the way of my process.
Way too rigid, but about the same as GPT-Codex-Max. In other words, it's all business and doesn't try to have a conversation in a coding session. Some people like that. I think it's an unrealistic way to approach a paired coding session, personally.
And I feel like this particular common alignment headspace gets in the way, when I'm probing the agent to determine what the best agent harness might look like on any given application. I think the words I use around this are triggering alignment safeguards and self denial, and the alignment mechanism isn't thinking as deeply as context. That's a problem, but it's a common one. This is why we test.
But in terms of output, as good or better than Codex high, and better than Sonnet 4.5.
The output is well organized, it follows my coding standards to the letter, and code quality is exceptional for an AI. One thing I did notice, is that the agent wants to optimize design requirements, but doesn't insist on it.
It also jumps to conclusions without asking for clarification. It's annoying that they all do this. I don't want confidence when a simple clarifying question would save hours of work.
That said, it seems to be handling a non-trivial amount of complexity better than open weights models usually do.
I'm quite pleased with it.