Who this is for and what changed
If you already call open-weight models through Hugging Face, this guide covers what changed when Inference Providers moved language model requests to OpenRouter. Your code stays the same, but routing, billing and API keys work differently now. After reading it, you should know how to choose a routing policy, control which providers serve you, and check where your spending went.
How requests are routed now
Language model calls are now sent through OpenRouter. For each call, it weighs current price, speed and reliability to choose a provider, and if that provider fails, it tries another one.[1] Before the change, Inference Providers chose one provider for each request and returned that provider's error when the request failed.[1]
It accepts requests in the OpenAI Chat Completions and Responses formats and in Anthropic's Messages format, covering all chat models regardless of the backend provider.[1] Setup guides cover Claude Code, Codex, OpenCode, Pi and Hermes Agent.[1]
Choose a routing policy
Without a suffix, the router balances price, speed and reliability.[1] A suffix on the model ID changes that choice:[1]
zai-org/GLM-5.3 # default: best mix of price, speed and reliability
zai-org/GLM-5.3:fastest # highest throughput
zai-org/GLM-5.3:cheapest # lowest price
zai-org/GLM-5.3:novita # only Novita:fastest selects the highest throughput, and :cheapest selects the lowest price. A provider suffix such as :novita sends requests only to that provider.[1]
What happens behind the router
Each month OpenRouter handles over 800 trillion tokens for more than ten million people, spanning 600-plus models and 80-plus providers.[1] Its routing was rebuilt to favor the provider with the strongest combination of price, speed and reliability instead of the lowest list price.[1]
If a provider gets slow or unreliable, the router stops sending it traffic until it is healthy again. A failed attempt goes to the next-best provider in the same call.[1] In OpenRouter's measurements, the new routing cut capacity errors by 46%, improved time to first token by 16.9% and raised throughput by 16.3%.[1]
Keep a conversation on one provider
The router also takes caching into account, keeping a conversation on the provider that already holds its prompt cache.[1] To make this explicit, send a stable UUID for each conversation in the X-HF-Session-Id header.[1]
Pricing and credits
The price stays the same: charges for each provider show up on your Hugging Face invoice, with no extra margin added.[1] The monthly allowance is $2 of credit for PRO accounts, and $2 per seat for Team and Enterprise organizations.[1]
Check which provider served you
A model page's Providers tab shows each provider that can serve the model through OpenRouter, with input and output prices plus speed and uptime figures for the past day.[1] The Metrics tab summarizes the top figures across those providers: lowest price, peak throughput, quickest first token, context window and uptime.[1] Spending in the usage overview is grouped by provider, so you can see where each request ran.[1]
In settings, you can turn off any provider. The router never sends requests to a provider you have turned off. Organizations manage the same switches, which then cover all requests charged to that organization.[1]
Own API keys no longer apply to these models
This is a breaking change. Charges for language model calls land on your Hugging Face account or organization, which means a provider key of your own no longer works with these models.[1] A model:provider suffix still selects a provider, and disabling providers still works. No code edits are needed.[1] Keys you bring for image and video providers still function.[1]
Image, video and audio are unchanged
Media generation for images, video and audio does not change: those requests skip OpenRouter and go directly through Hugging Face Inference Providers.[1] The same Hugging Face token also works with Storage Buckets, Jobs and Inference Endpoints.[1]
Why open-weight models are getting more traffic
On average, open-weight models cost about four times less: input tokens run $0.28 per million and output $1.62, while closed models cost $1.27 and $6.36.[1] Measured on OpenRouter, the open-weight share of tokens went from 26.8% in October 2025 up to 71.7% by late September 2026.[1]
Next steps
Connecting an OpenRouter account to Hugging Face is the next step in the partnership, and more tasks are planned to route through OpenRouter.[1]
Sources
- 1
Cite this page: Running open-weight models on Hugging Face through OpenRouter. TellWorld, https://www.tellworld.net/practice/practice-hf-inference-openrouter. Updated 2026-11-18