Hugging Face Moves Inference to OpenRouter
Hugging Face on Nov. 16 replaced its in-house router for Inference Providers with OpenRouter's, which picks a provider per call by live price, speed and reliability and retries on failure. Pricing is unchanged, but custom provider keys no longer work for language models.
Hugging Face on Monday, Nov. 16, switched the request router behind its Inference Providers service from its own in-house system to OpenRouter's, sending every call to a hosted language model through the third-party gateway. The company said developers' existing code and Hugging Face tokens keep working, but users who brought their own API keys for language-model providers lose that option.

Key takeaways
- Chat-model requests on Hugging Face Inference Providers now run through OpenRouter, which picks a provider per call by live price, speed and reliability and retries on another provider if one fails.
- Pricing is unchanged and Hugging Face adds no markup, but custom provider keys no longer work for language models.
- OpenRouter says 160 models from 38 labs are available through the integration at launch.
How Hugging Face's new OpenRouter routing works
In its announcement, Hugging Face said requests to language models "now go through OpenRouter, which picks the best provider for each call based on live price, speed and reliability, and retries on another provider if the first one fails." Previously, Inference Providers chose one provider and returned that provider's error if the call failed.
Developers can still steer the choice by adding a suffix to the model ID: “:fastest” for the highest throughput, “:cheapest” for the lowest price, or a provider name to pin a single host. Routing is cache-aware, so a conversation stays on the provider holding its prompt cache. The router accepts OpenAI's Chat Completions and Responses APIs as well as Anthropic's Messages API.
Image, video and audio generation are not affected. Those requests continue to run directly through Hugging Face, and custom keys for image and video providers keep working.
What changes for developers' bills and keys
Hugging Face said users pay the provider's price on their Hugging Face bill with no markup. PRO users get $2 in monthly credits, and Team and Enterprise organizations get $2 per seat.
The one breaking change is bring-your-own-key access. Because language-model calls are now billed to the Hugging Face account, users "can no longer use your own API key for a provider with these models," the company wrote. Users can still switch off any provider in their settings, and organizations can apply that rule to every request billed to them. Model pages gain a Providers tab listing each host's price, speed and 24-hour uptime.
Why OpenRouter's routing numbers matter
OpenRouter says it serves more than 800 trillion tokens a month to over 10 million users, across 600-plus models from more than 80 providers. In an Oct. 28 post, it said it had replaced mostly price-based routing with a score that weighs effective price, throughput, latency, error rates and output quality.
In a one-week switchback test covering more than 24 million requests, OpenRouter reported that the new router cut the share of requests hitting a rate-limit error by 46%, lowered median time to first token by 16.9% and raised median throughput by 16.3%. Cost per request rose 6.0%. Those figures are OpenRouter's own measurements and have not been independently verified.
OpenRouter's own announcement said the integration also runs the other way: in beta, it can send a JSON trace of each request, including prompts, outputs, token counts and cost, to a Hugging Face Storage Bucket the user owns, for building evaluation or fine-tuning sets.
Context: open-weight models now dominate gateway traffic
The deal pairs the main hub where open-weight models are published with a gateway where those models carry most of the volume. According to OpenRouter data published Oct. 31, open-weight models went from 26.8% of tokens among its top 50 models in the week of Oct. 6, 2025, to 71.7% in the week of Sept. 28, 2026, after peaking at 77.5% in July. Over the last four weeks of that period, token-weighted list prices averaged $0.28 per million input tokens for open-weight models versus $1.27 for closed models.
The shift mirrors earlier gateway data. AI Newsway has reported on how routing layers reorganized around Chinese open-weight releases.
What's next
Hugging Face called the switch "the first step" of the partnership. Next on its list are letting users connect an OpenRouter account to Hugging Face and routing more task types through OpenRouter. Neither company gave a date.
FAQ
Do I need to change my code?
No. Hugging Face said existing code and tokens keep working. Only developers who used their own provider API keys for language models need to change their setup.
Will Hugging Face inference cost more?
Hugging Face says list pricing is unchanged and it adds no markup. OpenRouter's own test found the new routing raised cost per request by 6.0% because it sometimes picks faster or more reliable providers over the cheapest. Adding “:cheapest” to a model ID favors the lowest price.
Can I block a specific provider?
Yes. Any provider can be turned off in Hugging Face's Inference Providers settings, and the router will not send requests there.
References
- Hugging Face and OpenRouter join forces to make open-weight models easy to run — Hugging Face (Nov. 16, 2026)
- OpenRouter Now Powers Inference on Hugging Face — OpenRouter (Nov. 16, 2026)
- How we rebuilt provider routing for better, faster inference — OpenRouter (Oct. 28, 2026)
- Open-Weight vs Closed-Weight Models: What's the Difference? — OpenRouter (Oct. 31, 2026)
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.