I am experiencing something similar. I’ve got the GLM-4.7 BYOK configured, and the token usage seems quite similar when I use GLM or Sonnet
Hey @ToX!
Happy to look into this. What endpoint are you using for BYOK?
Hello @Colin , I’m using this endpoint, as suggested in their guides: **https://api.z.ai/api/coding/paas/v4
**
Next time you use glm-4.7 with BYOK, could you grab a Request ID?
Hello Colin, here is a couple I’ve used today:
a971c91e-80a7-4eca-82e8-9864b9c7e6ca
aa6915d7-6ccc-430a-a18d-c35b835408d5
According to the stats, the cost was 0.60 USD
Thanks @ToX!
Our logs show that request as routing through Cursor’s API key (weird to think of it that way, but basically not-BYOK) and to Fireworks.
GLM-4.7 is supported “behind the scenes” in Cursor, and I wonder if somehow that fact is overriding your own settings.
Could you share a screenshot of the API Keys section of your Cursor Settings > Models?
Hey @ToX
I have the feeling that neither the OpenAI API Key nor the OpenAI Base URL is affecting your instance, and I’m losing faith that this has anything to do with how the model is ultimately routed once the request reaches us. I’ve sent similar values on my instance, and all calls to glm-4.7 fail (because my API key isn’t valid). Meanwhile, in your cases, your requests are never making it to z.ai.
Leaving those on, could you try using any other built-in model (Flash 3, Haiku 4.5, anything)? The expectation is that they would fail, because these overrides affect all model requests, not just BYOK models.
Toggling them on and again might kick them into place.
