From my testing it was ok until 145k tokens, the largest context I had before switching to a new session. I think Z.ai officially said it should be good until 200k tokens.
Using it in Open Code is compacting the context automatically when it gets too large.