Hugging Face Hands Its Inference Router to OpenRouter, and Bring-Your-Own Provider Keys Go Away
Starting November 16, 2026, chat-model requests on Hugging Face Inference Providers run through OpenRouter's router, which picks a provider per call on live price, speed and reliability and retries on failure. 160 models from 38 labs are covered, pricing carries no markup, custom provider keys no longer work for language models, and OpenRouter can now write request traces to Hugging Face Storage Buckets.
Hugging Face has handed language-model routing on its Inference Providers service to OpenRouter. On November 16, 2026, Hugging Face said it is replacing the custom router behind Inference Providers with OpenRouter’s. From that date, requests to chat models made with a Hugging Face token go through OpenRouter, which chooses a provider for each call.
Hugging Face’s announcement frames the change as a way to make open-weight models “easy to run.” Developers do not need to change their code: the same token, endpoints and model IDs keep working. What changes is how a provider is picked, what happens when one fails, and which keys are accepted.
How a request now gets served
Under the old system, Inference Providers picked one provider and returned its error if that provider failed. Under OpenRouter, each request goes to the provider with what both companies describe as the best mix of live price, speed and reliability. If a provider slows down or starts failing, traffic moves away from it, and a failed attempt is retried on the next provider within the same call.
OpenRouter says it rebuilt this routing in October and ran a live experiment. It reports 46% fewer requests hitting capacity errors, 16.9% faster time to first token and 16.3% higher throughput, at 6.0% higher cost per request.
Developers can still steer the router with a suffix on the model ID. Hugging Face’s examples use zai-org/GLM-5.3: the bare ID uses the default policy, :fastest favours throughput, :cheapest favours price, and a provider name such as :novita pins the call to one host. Routing is also cache-aware, so a conversation stays with the provider holding its prompt cache. Developers can make that explicit by sending a stable per-conversation UUID in an X-HF-Session-Id header.
The router accepts OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages API for every chat model. Hugging Face publishes setup guides for Claude Code, Codex, OpenCode, Pi and Hermes Agent.
Coverage and price
OpenRouter says 160 models from 38 labs are available this way at launch, which it describes as every model it serves that also has a page on the Hub. The list it highlights includes DeepSeek V4.1 Flash, Kimi K3.1 Pro and Kimi K3, GLM 5.3 and GLM 5.3 Flash, Qwen3.8 27B, MiniMax M3.1 Flash, Xiaomi MiMo-V2.6-Flash, NVIDIA Nemotron 3 Ultra and OpenAI’s gpt-oss-120b.
Pricing does not change. Users pay the provider’s price on their Hugging Face bill, and Hugging Face says it adds no markup. PRO users keep $2 in monthly credits, and Team and Enterprise organisations get $2 per seat.
Each model page gains two tabs. Providers lists every host OpenRouter can route to, with price, speed and 24-hour uptime. Metrics shows the best price, throughput, time to first token, context length and uptime among them. Usage pages now break spend down by provider, and users or organisations can switch off any provider they do not want to receive their traffic.
The breaking change
One thing does stop working. Because language-model requests are now billed through Hugging Face, developers can no longer use their own API key for a provider with those models. The company’s guidance is to use the model:provider suffix or provider settings instead. Custom keys for image and video providers still work, and image, video and audio generation keep running through Hugging Face Inference Providers directly, without OpenRouter.
Traces flowing back to the Hub
The integration also works in the other direction. OpenRouter can now send a trace of each request to a Hugging Face Storage Bucket owned by the user, through its Broadcast feature, which is in beta. Each trace is a JSON file holding input messages, model output, token counts and cost; images and other media are left out. Files can be organised by date or API key. OpenRouter pitches this as raw material for evaluation sets or for fine-tuning an open-weight model on real traffic.
Open-weight models now carry most OpenRouter tokens
Both companies lean on usage data. Hugging Face cites OpenRouter figures showing open-weight models rising from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026, after peaking at 77.5% in July. It also says open-weight models cost about four times less on average, while acknowledging that the best of them still trail top proprietary models on the hardest benchmarks.
Hugging Face calls this the first step of the partnership. Next on its list are connecting OpenRouter accounts to Hugging Face and routing more task types through OpenRouter.
- #Hugging Face
- #OpenRouter
- #Open-Weight Models
- #Inference
- #Model Routing
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.