AI Aulis

AI NEWS IN ENGLISH · 17 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 17 November 2026Paper

Hugging Face now routes language model requests through OpenRouter’s router

AI-generated contentWritten by AI

huggingface.co, x.com, technologyreview.com · 4 sources · Sources published: 16 November 2026

Hugging Face’s Inference Providers now passes language model requests to OpenRouter’s router, which picks a provider by price, speed and reliability and retries a failed request with another provider. Providers’ own API keys no longer work with language models.

Hugging Face announced the change on its blog on Monday 16 November. Inference Providers is a Hugging Face service that lets users call models published on the Hub with a Hugging Face access token, even when an outside provider runs the model. Until now, the service used its own router. Users do not need to change their code, Hugging Face said.

Failed requests move to another provider

For each request, OpenRouter’s router picks the provider with the best combination of price, speed and reliability. It bases the choice on OpenRouter’s continuous measurements of more than 80 providers.

Previously, Inference Providers returned a provider’s error to the user. Now a failed request is retried with the next-best provider during the same call. Users can still ask for the fastest, the cheapest or a specific provider by adding a suffix to the model ID (:fastest, :cheapest or the provider name, such as :novita). Requests in the same conversation stay with the provider whose cache holds the conversation’s prompt.

OpenRouter recently overhauled its own routing, which previously looked only at the lowest list price. In OpenRouter’s own measurements, the overhaul cut capacity errors by 46 per cent and improved time to first token by 16.9 per cent and throughput by 16.3 per cent. The figures therefore do not show the effect of Hugging Face switching routers.

For all chat models, the router supports OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages API. Image, video and audio generation work as before, directly through Inference Providers without OpenRouter.

Providers’ own keys no longer work with language models

Language model requests are now billed to the user’s Hugging Face account or organisation, so users can no longer use a provider’s own API key with language models. Users can still choose a provider with a model ID suffix, or block it in their settings so that the router sends it no requests. Image and video providers’ own keys still work.

Pricing does not change: users pay the provider’s price, and Hugging Face adds no markup. PRO subscribers get $2 of credits a month, and Team and Enterprise organisations get $2 per seat. New tabs on model pages show providers’ prices, speed and availability over the past 24 hours.

Hugging Face cites OpenRouter’s figures to justify the partnership: open-weight models’ share of the tokens OpenRouter routed rose from 26.8 per cent in October 2025 to 71.7 per cent at the end of September 2026. The company admits that the best open-weight models still trail the best closed models on the hardest benchmarks. On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro scored 46 points and Claude Fable 5.2 scored 65, but Kimi cost about a seventh as much per task.

Next, Hugging Face plans to let users link an OpenRouter account to Hugging Face and to route other tasks through OpenRouter too. On X, Hugging Face’s Clement Delangue also mentions access to model weights directly from OpenRouter and storing OpenRouter agent traces on Hugging Face. The image in Delangue’s post shows an OpenRouter model page linking to the model weights, but the blog does not describe storing traces.

The blog gives setup instructions for Claude Code and Codex, among others, and the Inference Providers documentation has further details.

Report an error in this story