Hugging Face and OpenRouter team up: language-model requests on Hugging Face now run through OpenRouter's router
IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds
Hugging Face is handing OpenRouter the job of dispatching calls to language models made through its Inference Providers service. OpenRouter picks, out of more than 80 providers, the one that best combines cost, speed and reliability, and tries another when the chosen one errors. Hugging Face keeps the rate free of markup and asks for no code changes, though personal provider keys stop working with these models.
Hugging Face and OpenRouter, the service that spreads requests across model providers, have partnered so that open-weight models are simple to run. The first effect lands today: Inference Providers, the Hugging Face service for using models through outside providers, is retiring its in-house router and adopting OpenRouter's for language models, the company explains on its blog.
After the switch, each request is sent to whichever provider currently offers the best blend of price, speed and reliability, out of more than 80. When that provider fails, the request is tried again on a different one within the same call; previously, Inference Providers chose a single provider and passed back its error. Anyone who prefers to can still set a policy or a specific provider by adding a suffix to the model ID (fastest, cheapest or a named one) and switch off unwanted providers in the account settings.
User code stays as it is, and so does the price: customers pay what the provider charges, on the Hugging Face bill and with no markup. The trade-off is that personal provider API keys no longer work with language models, because those requests are now billed to the Hugging Face account. Image, video and audio generation carry on as before, without passing through OpenRouter, and personal keys for image and video providers remain valid there.
Hugging Face describes the switch as the first step of the partnership. Next, it says, will come linking an OpenRouter account to a Hugging Face one and routing more kinds of tasks through OpenRouter. According to figures cited in the post, OpenRouter handles upwards of 800 trillion tokens every month for more than 10 million users.
Why it matters · analysis and opinion
Hugging Face gives up maintaining a component of its own and leans on the player that already moves this traffic at scale, a sign that spreading requests across providers has become a trade in itself. For developers the gain is practical: fewer visible errors when a provider goes down, with no code changes and no extra cost. The price of that convenience is depending on one more intermediary and losing personal keys, the way to use the account each team already had with a specific provider. Anyone in that position should review their language-model integrations today, because the change is already in force. It is also worth deciding which providers to switch off in the settings, especially where there are requirements about where data is processed.
Official source: Hugging Face · Written with the help of AI: how we make the news
What does it mean for your business? · analysis and opinion
This change affects small businesses, development studios and freelancers in Spain and Latin America that already call open-weight models from their applications using Inference Providers, the Hugging Face service for using models through external providers. Anyone who only uses ChatGPT, Gemini or Claude from their website will not notice anything, and neither will anyone who generates images, video or audio with Hugging Face, which keeps working the same way.
The urgent task is to check whether any integration used its own provider API key with language models: that stops working, and everything is now billed to the Hugging Face account or the organization. Our advice is to go into the settings and turn off any providers you do not want to share data with, because the router will never send requests there, and in an organization that setting applies to everything billed to it. If budget is the priority, it is enough to add the suffix :cheapest to the model identifier to request the lowest price, or to set a specific provider by name.
In terms of cost, Hugging Face does not add a markup on top of the price charged by each provider, and PRO accounts get $2 in credits every month, the same as each seat on the Team and Enterprise plans. It is wise to check the usage overview during the first few weeks, which now breaks down spending by provider, to confirm that the bill and the destination of the data match expectations.


