Solved. Don't mind this.

[semanticksz]

Solved!

[diegosouzapw]

Thanks @semanticksz. This is the same failure signature that was reported in #11729 on 3.8.50 (still open): the Z.ai page opens a chat, but the request our browser transport waits for (/api/v2/chat/completions) never fires, only POST /api/v1/chats/new does. So the web transport didn't get to the actual chat call.

A change aimed at this area went out in v3.8.51, published yesterday: #12524, "refresh Z.ai Web models and browser transport". I haven't validated it against the live service, so I can't tell you it fixes this. To see where you stand, I need:

1. Your OmniRoute version (omniroute --version, or the image digest for Docker). If you're still on 3.8.50 or an older :next, update to 3.8.51 first and retry.
2. Whether it still fails on 3.8.51, and with which exact model id (you wrote "GLM 5.3 flash").
3. If it does, the app.log lines around one failure (tokens and cookies redacted), and whether another Z.ai model such as a GLM 5.x non-flash one works on the same connection.

If it still fails on 3.8.51, your report becomes new data on #11729 and I'll add it there. Qwen Web, which you mentioned, was retired on purpose (#11713), so that one isn't a bug.