Hugging Face has switched language model requests from self-hosted routing to OpenRouter: it selects providers based on real-time pricing, speed, and reliability, and reroutes requests within the same call if one fails—all without requiring changes to existing code.

The model entry point hasn’t changed, and the routing layer now handles compute procurement by default. The trade-off is that language models no longer support bring-your-own provider keys. If the actual error rate and latency aren’t consistently lower than with the old routing, this middleware has no value.