Hugging Face Inference Providers now route via OpenRouter
On November 16, 2026, Hugging Face began routing chat models through OpenRouter. Your own provider API keys no longer work for language models.

On November 16, 2026, Hugging Face began routing chat models through OpenRouter. Your own provider API keys no longer work for language models.

On November 16, 2026, Hugging Face published Hugging Face and OpenRouter join forces to make open-weight models easy to run ↗. OpenRouter published OpenRouter Now Powers Inference on Hugging Face ↗ the same day. Tensor Journal ↗ reported the switch. The picture above is the announcement image on the Hugging Face post. The marks are uncropped.
Hugging Face says it switched Inference Providers for language models from its own router to OpenRouter's. You keep the same Hugging Face token, and the post says your code stays the same, through the same API. The routing docs ↗ say the Hugging Face router still authenticates and bills the call, then OpenRouter picks the provider. Every chat model accepts OpenAI's Chat Completions API, OpenAI's Responses API, and Anthropic's Messages API. Setup guides cover Claude Code, Codex, OpenCode, Pi, and Hermes Agent.
Without a suffix, OpenRouter picks a provider from the live effective price, speed, and reliability. The docs say effective price includes the real cache-hit rate and the usual mix of input and output tokens, not the list price alone. A failed try is retried on the next provider inside the same call. A pinned provider is never replaced. Pinning a provider that does not serve the model returns an error.
| Model id suffix | Where the call goes |
|---|---|
| none | Best mix of live price, speed, and reliability |
:fastest | Highest throughput |
:cheapest | Lowest price |
:novita, or another provider | That provider only |
Routing is cache-aware: a conversation stays on the provider that holds its prompt cache. To make that explicit, send a stable id per conversation in the X-HF-Session-Id header.
You pay the provider's price on the Hugging Face bill. Hugging Face does not add a markup. PRO accounts get $2 of credits a month. Team and Enterprise organizations get $2 per seat. Image, video, and audio generation stay on Hugging Face's own path, not through OpenRouter.
The breaking change: your own provider API keys no longer work for language models. Those requests bill your Hugging Face account or organization. You can still pin a provider with a suffix, or turn providers off. Custom keys for image and video providers still work. Organization settings apply to organization-billed requests. The providers docs ↗ say those requests follow the organization settings rather than yours.
Model pages gained Providers and Metrics tabs: price, speed, and uptime over the last 24 hours, plus a usage breakdown by provider.
OpenRouter's blog, on an October routing rebuild, reports 46% fewer capacity errors, 16.9% faster time to first token, and 16.3% higher throughput, for 6.0% higher cost per request. Hugging Face prints the first three figures and omits the cost. That 6.0% is OpenRouter's measurement, not a Hugging Face markup.
OpenRouter can send each request as a trace to a Hugging Face Storage Bucket you own (Broadcast, beta). The file holds messages, output, token counts, and cost. Images and other media are left out.
Hugging Face, citing OpenRouter, says open-weight models were 26.8% of OpenRouter tokens in October 2025, 71.7% at the end of September 2026, and peaked at 77.5% in July. Average open-weight price: $0.28 per million input tokens and $1.62 per million output tokens, against $1.27 and $6.36 for closed models. Next, Hugging Face says, connect an OpenRouter account, and route more tasks through OpenRouter.
"Each request goes to the provider with the best price, speed, and uptime for that model, with automatic failover." — OpenRouter ↗ (@OpenRouter), the company account
"Open-weight models went from 27% to 72% of tokens on OpenRouter in under a year." — Alex Atallah ↗ (@alexatallah), OpenRouter's CEO. The blog's figures are 26.8% and 71.7%.
"More than ever open source should win, and to get there, we also need to make open models easier to use for all!" — célina ↗ (@hanouticelina), a Hugging Face engineer
"If you had credits or a contract tied to one provider, check that before your next deploy." — The Lazy Edge ↗ (@TheLazyEdge)
"Same API does not always mean the same billing setup." — Nikhil Goyal ↗ (@nikhilgoyalai)
Developers who wanted one Hugging Face token and a failover win the default. The call stays a chat-completions or Messages request. The router moves a failed try to the next provider, unless you pinned one.
Anyone with a provider key on a language model has to act before the next deploy. Those calls now bill the Hugging Face account. A credit balance or a contract at a single host does not pay this route. Image and video keys are the exception the post names. If a compliance rule forbids a provider, turn it off in settings. Organization-billed traffic follows the organization settings.
The same model is three bills. Hugging Face: one token, failover, provider price, no Hugging Face markup. The 6.0% figure is OpenRouter's measurement of its October experiment, not a Hugging Face fee. Calling OpenRouter directly is that router, part of Stripe since the November 10 close, on its own account. Calling a provider directly keeps that contract and drops this failover. Auto Code picks a coding model. Decider v1.3 is a decision checkpoint. DeepSeek V4.1 Flash Ultrafast is a Vercel speed tier. None of those is this Hub switch.
Language-model traffic on Inference Providers now depends on one router. The open-weight share and the $0.28 / $1.62 averages are Hugging Face and OpenRouter's figures. Weights you run yourself are a different path, in the open-source LLM comparison and the open-source AI category.
Leave the default on if you already bill Hugging Face and want failover. Pause if the language-model budget still sits on a provider invoice, or if you must name the company that saw the prompt.