Hugging Face now routes all language-model requests on its Inference Providers service through OpenRouter, a change announced on 16 November, the company said in a blog post. OpenRouter picks one of more than 80 providers for each request based on price, speed and reliability, and open-weight models there cost about a quarter as much as closed ones on average, according to the post.
Key facts
- Since 16 November 2026, language-model requests on Hugging Face Inference Providers go through OpenRouter's router; image, video and audio stay with Hugging Face.
- Open-weight models carried 71.7% of OpenRouter tokens at the end of September 2026, up from 26.8% in October 2025, according to Hugging Face.
- Customers can no longer use their own provider API keys for language models; Hugging Face says it adds no markup to provider prices.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
The change covers language models only. Image, video and audio generation still run directly through Hugging Face's own Inference Providers, the company said. Existing code and Hugging Face tokens keep working unchanged, according to the post. The router supports OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API, with setup guides for Claude Code, Codex, OpenCode, Pi and Hermes Agent.
If a provider fails, the router retries the next one within the same call, Hugging Face said. A suffix on the model name changes that: “:cheapest” takes the lowest price, “:fastest” the highest throughput, and a provider name pins the request. Users and organisations can switch off individual providers in their settings.
Pricing and the end of custom provider keys
Hugging Face says it adds no markup to the provider's price. PRO users get $2 of credits a month, and Team and Enterprise organisations get $2 per seat. Customers can no longer use their own provider API keys for language models; billing now runs through the Hugging Face account, the company said. Custom keys for image and video providers still work.
In a live experiment in October, the new routing cut requests hitting capacity errors by 46%, improved time to first token by 16.9% and raised throughput by 16.3%, according to OpenRouter. Cost per request rose 6.0% in the same test, OpenRouter's blog says.
Open-weight models carry 71.7% of OpenRouter tokens
Hugging Face backs the move with OpenRouter data. Open-weight models rose from 26.8% of tokens served on OpenRouter in October 2025 to 71.7% at the end of September 2026, after peaking at 77.5% in July, the post says. Open-weight models cost $0.28 per million input tokens and $1.62 per million output tokens on average, against $1.27 and $6.36 for closed models.
The post is open about the performance gap. On the Artificial Analysis Intelligence Index, the best open-weight model, Kimi K3.1 Pro, scores 46 against 65 for Claude Fable 5.2, but costs $3.68 per task against $26.54, about a seventh. OpenRouter itself serves more than 800 trillion tokens a month to over 10 million users across more than 600 models from more than 80 providers, according to Hugging Face.
Analysis
When an open model costs a seventh as much for 46 points instead of 65, the routing of each request becomes a cost decision in its own right. On Hugging Face, OpenRouter's router now makes that call by default, weighting criteria it chooses. OpenRouter's own October test, in which errors fell but cost per request rose 6%, shows that the default is a trade-off made on the customer's behalf.
For marketing and product teams running copy, summaries or classification at volume, open-weight models make a strong case; Microsoft makes a similar price argument with MAI Code 2. A benchmark is no substitute for in-house measurement, as the comparison of Fable 5.2 and GPT-6.2 Astra on real work shows. The sensible setup is a small test set of real tasks, scored as cost per usable result as with Qwen4-Max, and a deliberately chosen provider or “:cheapest” instead of the default. Teams that used their own provider contracts through Hugging Face will need to redo the maths.
Update, 18 November: The US administration is weighing a disclosure rule for Chinese AI models such as Qwen and Kimi, Reuters reports.
Got AI news? Submit news →
FAQ
What changes on Hugging Face Inference Providers?
Since 16 November 2026, OpenRouter routes language-model requests across 80+ providers by price, speed and reliability. Code and tokens stay the same, Hugging Face says, but custom provider keys no longer work for language models.
How much cheaper are open-weight models?
According to Hugging Face, open-weight models on OpenRouter average $0.28/$1.62 per million input/output tokens, against $1.27/$6.36 for closed models, roughly a quarter.
Can I choose the provider myself?
Yes. A suffix such as “:cheapest”, “:fastest” or a provider name on the model id controls the choice, and providers can be switched off in settings.
Source
Hugging Face: Hugging Face and OpenRouter join forces to make open-weight models easy to run
More sources
- OpenRouter – OpenRouter-Blog vom 16.11.2026: Live-Versuch Oktober (−46 % Kapazitätsfehler, +16,9 % TTFT, +16,3 % Durchsatz, +6,0 % Kosten pro Anfrage).




