Hugging Face Hands Your Chat Requests to OpenRouter and Shuts Out Your Own Provider Keys

Hugging Face switched language-model routing on Inference Providers to OpenRouter, promising the same token and code with automatic failover. The catch is that bring-your-own provider keys no longer work for chat models.

Nov 17, 2026
Hugging Face Hands Your Chat Requests to OpenRouter and Shuts Out Your Own Provider Keys
A crashed screen shows an error among Heathrow Airport departure boards

Hugging Face just handed one of its most important developer jobs to someone else. In a blog post dated November 16, the company said it is switching Inference Providers from its own custom router to OpenRouter's. Every request to a language model now goes through OpenRouter, which picks a provider for each call based on live price, speed and reliability, and retries on another provider if the first one fails.

The pitch is that nothing breaks. Hugging Face says your code stays the same, and OpenRouter says existing Hugging Face tokens keep working with nothing to set up. For most people calling an open-weight chat model, that is probably true. For some, it is not.

The catch for bring-your-own-key users

Buried near the bottom of the announcement is a section Hugging Face labels a breaking change. Because language-model requests are now routed by OpenRouter and billed to your Hugging Face account or organization, you can no longer attach your own API key for a provider when calling these models. Anyone routing a negotiated rate or a private provider deal through Hugging Face just lost that option for chat traffic.

"Requests to language models now go through OpenRouter and are billed to your Hugging Face account or organization, so you can no longer use your own API key for a provider with these models." Hugging Face

Hugging Face points to two escape hatches. You can still pin a provider by adding a suffix to the model id, such as :novita, and you can switch off any provider you do not want in your settings. Custom keys for image and video providers keep working, because those tasks are not part of the switch.

Image, video and audio generation continue to run directly through Hugging Face Inference Providers, the company says, without touching OpenRouter. This is a text routing change, not Hugging Face outsourcing its whole inference business.

What do developers get for giving up control? Mostly resilience. Hugging Face concedes that before the switch, Inference Providers picked one provider and simply returned its error if that provider failed. Now a failed attempt is retried on the next best provider inside the same call. OpenRouter says it rebuilt this routing in October so it no longer chases the lowest list price alone, and in its own announcement it reports that a live experiment measured 46% fewer requests hitting a capacity error, 16.9% faster time to first token and 16.3% higher throughput.

Those are OpenRouter's numbers from its own traffic, not an independent test of the Hugging Face switch. And the same sentence carries a detail Hugging Face's post leaves out: OpenRouter says that experiment also came with 6.0% higher cost per request. Hugging Face insists pricing does not change, meaning you pay the provider's price on your Hugging Face bill with no markup on top. Both can be true, since the default router may favor a faster, pricier provider. That is the trade-off a middleman now makes for you unless you add :cheapest or :fastest to the model id.

Some touches are sensible. Routing is cache-aware, so a conversation stays on the provider holding its prompt cache, and a new Providers tab on model pages shows each provider's price, speed and uptime. Hugging Face's docs add that a pinned provider is never swapped out on failure, which means pinning buys predictability at the cost of the very failover the switch is selling.

The scale argument comes from OpenRouter. Hugging Face repeats OpenRouter's claim that it serves more than 800 trillion tokens a month to over 10 million users across 600+ models from 80+ providers. OpenRouter says 160 models from 38 labs are available through the integration today, including DeepSeek V4.1 Flash and OpenAI's gpt-oss-120b.

Hugging Face calls this the first step of the partnership. Its stated plans include letting users connect an OpenRouter account to Hugging Face and routing more tasks through OpenRouter. That second item is worth watching. If image, video and audio follow, the bring-your-own-key crowd could lose its last foothold on the platform, and the Hub, where Hugging Face says every major open-weight release already lands, would lean even harder on someone else's router to decide where those models run.