What Is Model Downgrading, and How Self-Hosting Fixes It

The short answer

"Downgrading" is when you think you are calling a flagship model, but a relay quietly routes the request to a cheaper, smaller or quantized model — quality drops and you cannot tell.

How it happens

• Relays profit on the token spread, so cheaper backends mean higher margin.
• At peak load they auto-fall-back to smaller models to cap cost.
• You still see the same model name, and still pay the flagship price.

How self-hosting fixes it

Connect directly to each vendor official API with your own accounts and keys. Requests reach the official endpoint point-to-point, with no layer in between that can swap the model. You get deterministic quality and the official price.

How to check

• Compare output stability between a relay and a direct connection on the same prompt.
• Verify the real model id and usage in the response body.
• Move critical paths to a direct self-hosted connection and lock quality in.