On November 16, 2026, Hugging Face handed OpenRouter the routing of its customers’ requests to language models, the kind of AI that produces text. Hugging Face is the platform where the major open models, AI models that anyone can download, are published. OpenRouter, a New York company, now picks the host that answers each of those requests. The host is the company whose servers compute the answer.
The immediate consequence: an API key obtained directly from a host no longer works for these models. An API key is the access code a program uses to reach a paid online service. Billing now goes through Hugging Face.
The change affects Inference Providers, the Hugging Face service for using these models over the internet, without installing them on in-house servers. OpenRouter says its own platform spreads the requests of more than 10 million users across more than 80 hosts.
OpenRouter picks the host for every request
Hugging Face launched Inference Providers on January 28, 2025. The service offers one web address and one API key for models running at partner hosts such as Together AI, Fireworks, Novita and the French company Mistral. The November 16 announcement, posted by seven members of the Hugging Face team, describes the old system. It picked a single host and, if the request failed, passed that host’s error back to the customer.
For each request, OpenRouter now compares the price, speed and reliability of the hosts serving the requested model. If the first host fails, the request moves on to another. Hugging Face says customers do not have to change their code.
Customers can steer that choice with a suffix, a short tag added to the end of the model name. The suffix “:fastest” asks for the quickest host, and “:cheapest” asks for the lowest price. A host’s name as the suffix limits requests to that host. An ongoing conversation stays with the host that already holds its earlier messages in a temporary memory called the cache.
OpenRouter says 160 models from 38 AI labs are available this way. The Hugging Face documentation lists 19 hosts for language models. Hugging Face says image, video and audio generation do not go through OpenRouter and work as before.
Billing moves to Hugging Face
Hugging Face flags one breaking change, meaning one that can stop existing setups from working. Requests to language models are now billed to the user’s Hugging Face account or to their organization’s. Keys issued by hosts now work only for image and video models. AlphaSignal, an AI newsletter, says teams that used their own Together or Fireworks keys for negotiated rates must now bill those requests through Hugging Face.
Hugging Face says it adds no markup to the host’s price. Subscribers to its paid PRO plan get $2 in credits a month. The Team and Enterprise plans give organizations $2 a month per seat, meaning per user. The pricing page says free accounts have to buy credits.
Every model page on Hugging Face gets two new tabs. “Providers” lists each host’s price, speed and uptime over the last 24 hours. “Metrics” shows the best figures among those hosts. Users can turn off a host in their settings, and organizations can do the same for every request they pay for.
The deal also works in the other direction. OpenRouter’s own customers can save a copy of each request to a storage space they own on Hugging Face. The copy includes the prompt (the text sent to the model) and the answer. The feature is in beta, a testing phase.
Open models lead on traffic and price but trail on the hardest tests
Hugging Face also quotes OpenRouter’s usage figures, measured in tokens. A token is the unit of text AI services bill for, roughly a fragment of a word. The figures compare open-weight models, whose weights are published and free to download, with closed models, whose weights are not. A model’s weights are the parameters it learns during training.
Open-weight models accounted for 26.8% of the tokens processed by the 50 most-used models on OpenRouter in October 2025. By the end of September 2026, their share stood at 71.7%, after a peak of 77.5% in July. Over the same period, their weekly token volume grew more than 60-fold. Volume for closed models grew almost tenfold.
Open models also cost less. In October 2026, a million tokens sent to an open model cost $0.28 on average, and a million tokens it produced cost $1.62. For closed models, the figures were $1.27 and $6.36, about four times as much. These are OpenRouter list prices, weighted by usage.
OpenRouter attributes the price gap to competition. Any host can run an open model. On October 21, for example, 31 hosts were offering the same model, DeepSeek V4.1 Flash.
Hugging Face acknowledges that the best open models still trail the best closed ones on the hardest tests. Its post relies on the index of Artificial Analysis, an organization that tests AI models.
As of November 16, the top-ranked open model on that index is Kimi K3.1 Pro, made by Moonshot AI, with 46 points. Claude Fable 5.2, a closed model from Anthropic, scores 65. Yet for each task tested, the open model costs about a seventh as much as the Anthropic model.
OpenRouter’s own figures and unclear rules on customer data
The usage and price figures come from OpenRouter, a party to the deal. Its method for usage shares, laid out in an October 31 post, counts only the 50 most-used models each week. It leaves out all other models, as well as unannounced ones that OpenRouter calls “stealth” models. OpenRouter says the stealth models reached 38.7 trillion tokens in the most recent week, a third of the 116.5 trillion counted in the week of September 28.
Hugging Face also cites the gains OpenRouter reports for its routing system, rebuilt in October. Compared with the previous system, errors caused by a host running out of capacity dropped 46%. The wait for the first word of an answer fell 16.9%, and the amount of text produced per second rose 16.3%.
OpenRouter’s own announcement adds a point the Hugging Face post leaves out. The gains were measured in a one-week experiment on live traffic, and the cost per request rose 6.0% during that week. Customers who put price first can use the “:cheapest” suffix.
Where customer data goes is poorly documented. Hugging Face’s security page for Inference Providers, checked on November 17, does not mention OpenRouter. It says Hugging Face keeps neither the content of requests nor the answers. The company’s technical logs, which hold no user data, are kept for up to 30 days.
Beyond that, the page points to each host’s own rules and says each host is responsible for its own security. That security also depends on the software a host runs. One example is vLLM, open-source server software that runs an open model for many users at once. Since June, the vLLM project has published several security advisories about data leaking between users of the same server.
OpenRouter says it stores neither prompts nor answers without the customer’s consent, except content flagged as potential child sexual abuse material. Its data collection page lists a second exception: hosts labeled “Safety Retention.” At those hosts, OpenRouter or the host may keep prompts and answers to prevent abuse, without using them for training. A setting lets customers exclude those hosts.
OpenRouter keeps each request’s metadata, meaning technical data such as the number of tokens and the response time. It also classifies a small sample of requests, anonymously, to build its statistics.
Neither announcement says which OpenRouter privacy settings apply to requests coming from Hugging Face. OpenRouter’s direct customers can exclude hosts that reserve the right to train their models on the data they receive. On OpenRouter’s Business and Enterprise plans, they can also have the content of their requests processed in Europe. Neither the Hugging Face post nor its documentation says whether Hugging Face customers get those options.
What it means
The deal changes nothing for teams that run models on their own machines. They can still download the weights from the Hub, Hugging Face’s model catalog. The change affects teams that use models remotely through Hugging Face. Their requests now pass through three companies: Hugging Face for sign-in and billing, OpenRouter to choose the host, and the host that computes the answer.
These teams gain automatic switching to another host when one goes down, and they can see each host’s prices. They lose the option of paying a host directly with their own key.
Another open-model tool changed its terms this month. Ollama, open-source software that runs open models on a personal computer, dropped the free plan from its cloud, meaning its online service. Since November 9, free accounts there have to buy credits.
On September 3, 2026, Nvidia announced an agreement to acquire Hugging Face. Nvidia’s announcement promised that developers would choose “the clouds and inference service providers they want,” in other words, their hosts. The new system keeps that choice through suffixes and settings, and runs the billing through Hugging Face.
Hugging Face calls the deal a first step. Next, it plans to link OpenRouter accounts to Hugging Face accounts and to send other types of tasks to OpenRouter.
Two questions remain open. One is whether Hugging Face’s security page will be updated to describe OpenRouter’s role. The other is whether European customers will get a way to have their requests processed in Europe.
Sources
- Hugging Face and OpenRouter join forces to make open-weight models easy to run, Hugging Face, November 16, 2026
- OpenRouter Now Powers Inference on Hugging Face, OpenRouter, November 16, 2026
- Open-Weight vs Closed-Weight Models: What’s the Difference?, OpenRouter, October 31, 2026
- How OpenRouter rebuilt provider routing for better, faster inference, OpenRouter, October 28, 2026
- Security & Compliance (Inference Providers documentation), Hugging Face, retrieved November 17, 2026
- Data Collection, OpenRouter, retrieved November 17, 2026
- Hugging Face Drops Its Chat Router and Hands Control to OpenRouter, AlphaSignal, November 16, 2026
- NVIDIA to Acquire Hugging Face, NVIDIA, September 3, 2026
Header image: MACCUS chart, data from OpenRouter (post of October 31, 2026).



