Skip to content

Open models

Hugging Face sends language model calls via OpenRouter

5 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: openrouter.ai

On November 16, Hugging Face moved language-model calls in Inference Providers off its own router and onto OpenRouter’s. OpenRouter now picks among as many as 32 providers for a single model, choosing on live price, speed and reliability, and retries on another provider if the first one fails, Hugging Face said. If your team plugged its own provider API key into these models, that key no longer works for them, because every call is now billed to your Hugging Face account. Your code and the provider prices stay the same.

You can still steer the router with a suffix on the model id: zai-org/GLM-5.3:fastest for the highest throughput, :cheapest for the lowest price, :novita for Novita only. According to the updated docs, a pinned provider is never swapped out, so a pinned call gets no retry. In settings you can turn off any provider. The router then never sends your requests there, and pinning a provider you turned off returns an error. Organizations get the same settings, and theirs apply to every request billed to the organization. To keep a conversation on the provider that holds its prompt cache, send a stable UUID per conversation in the X-HF-Session-Id header.

Image, video and audio generation still run directly through Hugging Face, and custom keys for image and video providers keep working. You pay the provider’s price on your Hugging Face bill with no markup. PRO users get $2 in credits a month, and Team and Enterprise organizations get $2 per seat. The router that now carries this traffic belongs to a company Stripe agreed in August to buy for more than $7 billion, Bloomberg reported.

Defaults that moved

Before the switch, Inference Providers sent each request to one provider and returned that provider’s error if it failed. The blog lists only one breaking change, the custom keys. The docs update that shipped the same day shows more. The default used to be the :fastest policy, the highest throughput in tokens per second. Now it is OpenRouter’s blend of price, speed and reliability. The :preferred suffix, which followed your own provider order from settings, has been removed from the Inference Providers docs. :cheapest used to mean the lowest price per output token and is now described simply as the lowest price.

OpenRouter recently rebuilt that routing. In its own measurements, the rebuild cut capacity errors by 46% and improved time to first token by 16.9% and throughput by 16.3%. The same sentence of its post reports a 6.0% cost increase.

Why open weights

Open-weight models are models whose trained weights are published for anyone to run. On OpenRouter they made up 71.7% of tokens served at the end of September 2026, up from 26.8% in October 2025, after peaking at 77.5% in July. They cost about four times less on average: $0.28 per million input tokens and $1.62 per million output tokens, against $1.27 and $6.36 for closed models. The best of them still trail. On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro scores 46 against 65 for Claude Fable 5.2 at its Max setting, at $3.68 per task against $26.54.

Our analysis

What we checked

On November 16 we compared Hugging Face’s router catalogue against OpenRouter’s public listings for seven models, then read the docs change that shipped with the switch. We made no billed requests.

How we check stories

The no-markup claim holds for list prices. The catalogue is smaller than the blog’s “80+ providers” suggests, though: that figure is OpenRouter’s overall count, and Hugging Face’s router shows 57. On many models, failover has little or nothing to switch to.

  • Of 143 provider prices on Hugging Face, 131 match OpenRouter’s price for the same provider on both input and output. Nine have no OpenRouter listing for that model, three are lower on Hugging Face, and none is higher.
  • Hugging Face’s router lists 164 chat models. 59 of them have exactly one provider, and the median is two, so by our reading a failed call on those 59 has nowhere to be retried. The blog says the Providers tab lists every provider OpenRouter can route a model to. Yet for GLM-5.3, Hugging Face shows 32 providers against 36 on OpenRouter, and DeepSeek’s own endpoint for DeepSeek V4.1 Flash 1017 is missing.
  • One model id can mean very different builds. On GLM-5.3, in a snapshot at 21:19 UTC, input prices ran from $0.0699 to $1.68 per 1M tokens and output from $1.40 to $5.28. Twelve providers are labelled fp4, six fp8, three nvfp4 and eleven carry no label; those labels mark builds stored at lower numeric precision. The full-precision version is a separate id, zai-org/GLM-5.3-BF16, served only by zai-org. Novita, the blog’s pinning example, is labelled fp8. It is third-cheapest on output at $1.98, but 27th of 32 on throughput.

Seeing where calls land

Hugging Face says this is the first step of the partnership. Next on its list are connecting your OpenRouter account to Hugging Face and routing more tasks through OpenRouter. For now, each model page has a Providers tab with every listed provider’s input and output price, speed and 24-hour uptime, plus a Metrics tab with the best figures among them. Your usage overview splits your spend by the provider each request went to. If a provider you don’t want shows up there, turn it off in your settings.

ClaudeAnthropic

Anthropic plans Wednesday IPO at $1.4 trillion

“A clear sign that the pursuit of profit is winning,” The Atlantic writes of Anthropic’s $1.4 trillion IPO, from a company that warns its own strategy could end humanity. Would you buy into an AI lab that warns about its own product, or sit it out?

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.