Latest
AI Affairs, home

Tuesday 17 November 2026

Corporate

Hugging Face replaces its inference router with OpenRouter’s network

Developers can keep their existing code and Hugging Face token, but language-model requests will now be routed across providers and billed to their Hugging Face account.

Blue-lit server racks with rows of drive bays and indicator lights
Photo: panumas nikhomkhai via Pexels

Hugging Face said on 16 November 2026 that it had replaced its own Inference Providers router with OpenRouter’s system, which selects among more than 80 providers for language-model requests instead of sending each request to one provider. The announcement connects Hugging Face’s model-hosting Hub to OpenRouter’s inference network: developers can call a supported open-weight model with their existing Hugging Face token and API.

Key points

  • Hugging Face says OpenRouter now selects providers for language-model calls using live price, speed and reliability measurements.
  • Users pay the provider’s price on their Hugging Face bill, without a Hugging Face markup.
  • Open-weight models made up 71.7% of tokens served on OpenRouter at the end of September 2026, up from 26.8% in October 2025, according to Hugging Face.
  • OpenRouter’s measurements put capacity errors 46% lower under its rebuilt routing system, Hugging Face says.

OpenRouter takes over language-model routing

Hugging Face hosts AI models, datasets and applications on its platform. Its Inference Providers service lets developers run models through external providers. Under the partnership, requests to language models pass to OpenRouter, an API gateway that directs inference calls across providers. Hugging Face says existing application code does not need to change.

OpenRouter chooses a provider for each call using current measurements of price, speed and reliability, according to Hugging Face. If that provider fails, the router retries the request with another provider. Previously, Inference Providers returned the error from the single provider selected for the request. Developers can instead select the cheapest or fastest routing policy, or specify one provider in the model identifier.

Hugging Face says OpenRouter’s rebuilt routing system reduced capacity errors by 46% in OpenRouter’s measurements. Time to first token improved by 16.9% and throughput by 16.3% under the same measurements. According to Hugging Face, OpenRouter handles a monthly volume exceeding 800 trillion tokens, covering more than 600 models and a user base of over 10 million.

The router also accounts for prompt caches. Hugging Face says a conversation remains with the provider holding its cached prompt, and developers can identify that conversation by supplying a stable session identifier. The service supports OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages API for chat models regardless of which provider handles a call.

Hugging Face keeps provider pricing

Customers pay the provider’s price through their Hugging Face bill, and Hugging Face says it adds no markup. Its PRO plan includes $2 in monthly credits, while Team and Enterprise organisations receive $2 per seat. Usage records show spending by the provider that handled each request.

There is a change for developers who previously used their own provider API keys. Hugging Face says language-model calls now have to be billed to a Hugging Face account or organisation, so those calls can no longer use a customer’s separate provider key. Developers can still select a provider through the model identifier or exclude providers in their settings. Organisation-level exclusions apply to requests billed to that organisation.

Image, video and audio generation continue to run directly through Hugging Face Inference Providers rather than OpenRouter, the company says. Custom provider keys continue to work for image and video requests. Hugging Face says model pages now list available providers alongside their prices, speed and uptime over the previous 24 hours.

Open-weight tokens reach 71.7% of volume

Hugging Face says open-weight models accounted for 71.7% of tokens served on OpenRouter at the end of September 2026, compared with 26.8% in October 2025. Their weekly token volume grew more than 60-fold over that period, against almost tenfold growth for closed models. The open-weight share peaked at 77.5% in July.

On 16 November, Hugging Face identified DeepSeek V4.1 Flash 1017 as OpenRouter’s most-used model in its current weekly ranking, with 59 trillion tokens. For open-weight models, Hugging Face cited average charges per million tokens of $0.28 for input and $1.62 for output, compared with $1.27 for input and $6.36 for output on closed models.

Hugging Face says the strongest open-weight models still score below the strongest proprietary models on the hardest benchmarks. It cited an Artificial Analysis Intelligence Index score of 46 for Kimi K3.1 Pro, compared with 65 for Claude Fable 5.2, and a cost per task of $3.68 against $26.54.

Sources

Topics: Foundation models, Inference, Open source

What should we cover? Tell us

The weekly briefing

The AI Affairs weekly briefing: the week's 10 most important AI stories in business, government and markets, every Monday. We'll email you a link to confirm. Every email has a one-click unsubscribe. Privacy policy