Hugging Face switches language model routing to OpenRouter

Hugging Face announced on November 16, 2026 that it is delegating language-model routing to OpenRouter. From now on, requests to chat models use OpenRouter’s routing layer, which selects providers based on live measurements of price, speed and reliability and retries on an alternate provider if the first attempt fails. The change keeps existing code and Hugging Face tokens unchanged for developers.

Routing, APIs and provider policies

OpenRouter’s router supports OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages API for chat models, regardless of the underlying provider. Hugging Face said the same token also continues to grant access to Storage Buckets, Jobs and Inference Endpoints, and that PRO, Team and Enterprise plans include monthly compute and inference credits.

Routing decisions are now based on a mix of price, speed and reliability rather than list price alone. OpenRouter reported that its rebuilt routing reduced capacity errors by 46% and improved time to first token by 16.9% and throughput by 16.3% in its measurements. Providers can be selected or excluded via a model-id suffix such as zai-org/GLM-5.3:fastest or :cheapest, and routing is cache-aware when a stable UUID is sent in the X-HF-Session-Id header.

Usage figures, costs and limits

OpenRouter reported that open-weight models rose from 26.8% of tokens in October 2025 to 71.7% at the end of September 2026, peaking at 77.5% in July. Weekly open-weight volume grew more than 60-fold over the period, compared with almost tenfold for closed models. The most used model this week on OpenRouter is DeepSeek V4.1 Flash 1017 with 59.2 trillion tokens, OpenRouter’s rankings show.

Reported average prices for open-weight models are about $0.28 per million input tokens and $1.62 per million output tokens, versus $1.27 and $6.36 for closed models. On the Artificial Analysis Intelligence Index, the best open-weight model Kimi K3.1 Pro scored 46 versus 65 for Claude Fable 5.2; cost per task was $3.68 for Kimi K3.1 Pro and $26.54 for Claude Fable 5.2, the publisher reported.

Pricing remains billed at the provider’s rate on the Hugging Face bill with no additional markup. Image, video and audio generation continue to route through Hugging Face Inference Providers outside OpenRouter. A breaking change: custom provider API keys can no longer be used for language models; billing is tied to the Hugging Face account or organization, though provider suffixes and provider opt-outs remain available.


Original source: Hugging Face Blog

Leave a Comment