Hugging Face switches its inference router to OpenRouter
Language-model requests on Hugging Face Inference Providers now go through OpenRouter, which retries failed calls on other providers, but users lose the option to bring their own provider keys.

- What happened
- On 16 November, Hugging Face replaced its own Inference Providers router with OpenRouter's for language-model requests.
- What’s new
- Failed requests are now retried on another provider, routing weighs live price, speed and reliability, and custom provider API keys no longer work for language models.
- Why it matters
- Developers calling open-weight models through Hugging Face should see fewer failed requests, but anyone relying on their own provider keys must change how they are billed.
Hugging Face said on 16 November that its Inference Providers service now sends language-model requests through OpenRouter instead of its own custom router. Developers keep the same Hugging Face token and API, and Hugging Face says existing code does not need to change. According to The New Stack, the integration went live covering 160 models from 38 AI labs.
What changes for developers
Under the old system, Hugging Face picked one provider per request and passed back that provider's error if the call failed. OpenRouter instead chooses a provider for each call based on live measurements of price, speed and reliability, and retries a failed attempt elsewhere within the same call. Developers can still steer routing by adding a suffix to the model id: one for the fastest provider, one for the cheapest, or a named provider. Routing also keeps each conversation with whichever provider is holding its prompt cache. The router supports OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API.
Hugging Face says pricing is unchanged: users pay the provider's price on their Hugging Face bill with no markup added. Model pages gain Providers and Metrics tabs showing each provider's price, speed and uptime, and users and organisations can block providers in their settings.
The announcement includes one breaking change. Because language-model requests are now billed through Hugging Face, developers can no longer use their own API keys for providers serving these models. Users' own keys for image and video providers are unaffected. Image, video and audio generation continue to run on Hugging Face's own infrastructure rather than through OpenRouter.
The routing trade-off
Hugging Face cites OpenRouter's measurements for its rebuilt routing: 46% fewer capacity errors, a 16.9% improvement in time to first token and a 16.3% gain in throughput. The New Stack reports a cost that Hugging Face's post does not mention: OpenRouter's October overhaul raised the average price per request by 6%. The New Stack notes that this figure relates to the October changes on OpenRouter rather than to the Hugging Face integration specifically.
Open-weight demand and ownership
Hugging Face frames the deal around rising use of open-weight models. It says their share of tokens served on OpenRouter rose from 26.8% in October 2025 to 71.7% by late September 2026. The New Stack, citing the same OpenRouter analysis, describes the 71.7% figure as a share of tokens processed by OpenRouter's 50 most-used models rather than of all traffic. Hugging Face concedes that the best open-weight models still lag top proprietary ones on the hardest benchmarks, but says they are roughly four times less expensive on average.
According to The New Stack, the partnership was announced within a week of Stripe closing its purchase of OpenRouter. Hugging Face calls the switch a first step. It plans to let users link their OpenRouter accounts and to route more tasks through OpenRouter, but The New Stack says no timeline has been given.