Hugging Face hands language-model routing to OpenRouter
Inference Providers now route every chat model through OpenRouter for lower latency and automatic failover, but bring-your-own-key billing for language models stops working.
Hugging Face switched the routing behind Inference Providers to OpenRouter on 16 November, for every language model regardless of which provider actually serves it. OpenRouter already handles more than 800 trillion tokens a month for over 10 million users, across 600+ models from 80+ providers, and it picks a backend per request on live price, speed and reliability, retrying on another provider automatically if the first one fails. Hugging Face says the switch cut capacity errors by 46%, improved time to first token by 16.9% and raised throughput by 16.3%.
The catch is billing. Bring-your-own-key no longer works for language models: requests now bill through your Hugging Face account at Hugging Face’s rate, even if you’d negotiated a cheaper deal directly with a provider like Together or Fireworks. Image and video models are unaffected, custom keys for those still work as before. Your existing code doesn’t change, and you can still steer routing with a suffix on the model id:
zai-org/GLM-5.3 # default routing
zai-org/GLM-5.3:fastest # highest throughput
zai-org/GLM-5.3:cheapest # lowest price
Why it matters: if you built on Hugging Face Inference Providers specifically to use your own negotiated provider rate for a chat model, that saving is gone as of this change, you’re paying Hugging Face’s price instead. Everyone else gets automatic failover and provider selection they didn’t have to build themselves.
The caveat: the performance numbers, 46% fewer capacity errors and a faster time to first token, are Hugging Face’s own, not independently verified.