Hugging Face switches its models to OpenRouter
Hugging Face transfers routing of its language models to OpenRouter, without changing the developer-side API. The duo promises better provider choices, with automatic switching based on price, speed and reliability.
Hugging Face now entrusts the routing of its language models to OpenRouter, according to an announcement published on November 16, 2026. Specifically, requests sent to the models go through a system that selects the most suitable provider for each call based on price, speed, and reliability. For developers, the API remains unchanged, but the underlying mechanism is significantly strengthened.
An invisible routing, but decisive for usage
OpenRouter already serves over 800 trillion tokens per month to more than 10 million users, relying on 600+ models and 80+ providers. Hugging Face adopts this component for its Inference Providers, with a simple logic: if a provider slows down or fails, the request is redirected to another. For language models, the service also adds a fallback attempt within the same request. Image, video, and audio models, however, continue to be served directly by Hugging Face without passing through OpenRouter.
The platform also highlights a cache-aware system: a conversation remains with the provider hosting its prompt cache, which limits unnecessary back-and-forth. Advanced users can still force a choice with a suffix on the model identifier, such as :fastest, :cheapest, or the name of a specific provider. Convenient. Nobody likes to see an API decide on its own, except when it does better than oneself.
What this alliance changes for developers
Hugging Face’s central message is continuity. OpenAI Chat Completions and Responses clients, as well as Anthropic’s Messages API, remain supported for compatible chatbots and agents. The same Hugging Face token also grants access to Storage Buckets, Jobs, and Inference Endpoints, while PRO, Team, and Enterprise plans include monthly credits for compute and inference.
However, a fundamental change appears: requests to language models are now billed to the Hugging Face account or organization, and it is no longer possible to use one’s own provider API key for these models. Hugging Face specifies that it does not take a margin on the provider’s price: the client pays the displayed rate, then routing handles the selection. According to OpenRouter, its new selection logic has reduced capacity errors by 46%, while improving time to first token by 16.9% and throughput by 16.3%.
Open models gain ground, for a very simple reason
On OpenRouter, open-weight models rose from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026, with a peak at 77.5% in July, according to figures published in October 2026. Over the same period, their weekly volume multiplied by more than 60, compared to nearly ten times for closed models. That week, the most used model on OpenRouter was also an open-weight: DeepSeek V4.1 Flash 1017, with 59.2 trillion tokens.
Price is a major factor. According to rates cited by OpenRouter, open models cost on average $0.28 per million input tokens and $1.62 per million output tokens, versus $1.27 and $6.36 for closed models. The best open models still lag on some difficult benchmarks, but the gap can become acceptable when the cost per task drops sharply. On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro scores 46 versus 65 for Claude Fable 5.2, for a cost per task of about $3.68 versus $26.54.
Key points
- OpenRouter serves over 800 trillion tokens per month.
- Open models represent 71.7% of tokens at the end of September 2026.
- Hugging Face routing now goes through OpenRouter.
- Capacity errors decrease by 46%, according to OpenRouter.
- DeepSeek V4.1 Flash 1017 is the most used this week.
- Hugging Face keeps the API but changes the engine behind it.
By the numbers
- 800+ trillion tokens per month — volume served by OpenRouter worldwide.
- 10 million users — base announced by OpenRouter.
- 46% — reduction in capacity errors after the new routing.
- 16.9% — improvement in time to first token, according to OpenRouter.
- 71.7% — share of open-weight tokens at the end of September 2026 on OpenRouter.