Hugging Face has moved all language-model traffic on its Inference Providers service onto OpenRouter’s router, retiring the custom router it built itself. According to the Hugging Face blog post announcing the change, one Hugging Face token now reaches any open-weight model on the Hub that a provider serves, and OpenRouter chooses the provider for every call. Teams that pay providers with their own keys lose that option for these models.

OpenRouter ranks providers by current price, speed and reliability, then retries on the next best one if a call fails. Before the switch, Inference Providers selected a single provider and passed back whatever error it returned. OpenRouter says it recently rebuilt its routing so it no longer leans only on the cheapest list price. By OpenRouter’s own count, capacity errors fell 46%, first-token latency improved 16.9% and throughput rose 16.3%. Those figures come from OpenRouter, the party being praised.

Existing code keeps working. The router accepts OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages API for every chat model, and Hugging Face published setup guides for Claude Code, Codex, OpenCode, Pi and Hermes Agent. A suffix on the model id steers the choice: :fastest for throughput, :cheapest for price, or a provider name such as :novita to pin one. Routing is cache-aware, so a conversation stays with the provider holding its prompt cache. A stable UUID in the X-HF-Session-Id header makes that explicit.

Pricing does not change. Customers pay each provider’s rate, shown on the Hugging Face invoice, with no markup added. PRO subscribers receive $2 in monthly credits, and Team and Enterprise accounts receive $2 per seat. Image, video and audio generation stay on direct Hugging Face routes, and custom keys for image and video providers still work.

The breaking change is the key one. Because requests now run through OpenRouter and are billed to a Hugging Face account or organization, a developer can no longer supply a personal provider key for language models. Users can still pick a provider by suffix, or switch providers off in settings. Organizations get the same controls, so an admin can block a provider for the whole account. Model pages gain Providers and Metrics tabs showing price, speed and 24-hour uptime, and the usage overview splits spend by provider.

The partnership rests on a shift in what developers run. OpenRouter reports that open-weight models went from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026, after a July peak of 77.5%. Weekly open-weight volume multiplied by more than 60, while closed-model volume grew nearly tenfold. The top model there this week is DeepSeek V4.1 Flash 1017, at 59.2 trillion tokens. OpenRouter overall handles above 800 trillion tokens monthly for 10 million-plus users, spanning 600+ models and 80+ providers.

Price explains much of that. Per million tokens, open-weight models average $0.28 in and $1.62 out. Closed ones average $1.27 in and $6.36 out. The post concedes a quality gap. On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro, the strongest open-weight entry, gets 46 while Claude Fable 5.2 gets 65. Cost per task is $3.68 versus $26.54. Hugging Face calls that trade-off worth it for many workloads. All of these usage and price numbers are OpenRouter’s, measured on OpenRouter’s own traffic.

The post calls this the first step. Next come connecting an OpenRouter account to Hugging Face and routing more tasks through OpenRouter. It does not say whether other task types will lose direct routes, or what Hugging Face receives from the arrangement.

Anyone running language models through Inference Providers on their own provider keys should check their billing setup now, because those calls move to Hugging Face invoices, and teams that need their own contract terms must go straight to the provider.

Hugging Face blog (huggingface.co/blog/openrouter), by Célina Hanouti, Victor Mustar and colleagues, with usage figures credited to OpenRouter, October 2026.