Symptoms
- Requests reach a website page, another provider, or a route that returns 404.
- Changing only base_url fixes every model.
Likely causes
- The OpenAI SDK base URL does not end in /v1.
- A full operation endpoint was pasted into a field that expects only a base URL.
How to diagnose
- Log the final request host and path with query strings and secrets removed.
- Identify whether the client field expects an origin, /v1 base URL, or full endpoint.
- Compare the final path with /v1/chat/completions or the required protocol endpoint.
Fixes
- Use the documented TokenReply URL shape for that client.
- For OpenAI SDKs, normally use the /v1 base URL; for raw HTTP, use the full operation path.
- Do not append a second provider hostname or operation path.
How to verify
The final request reaches the intended TokenReply API route and returns an API-shaped response.