Hugging Face announced on November 18, 2026 that its Inference Providers service will send all language‑model requests through OpenRouter. The change is automatic; existing API calls keep working without code changes. By default the router selects a provider that balances price, speed and reliability, but developers can add a suffix like ":fastest" or ":cheapest" to force a specific priority, or name a provider directly. A new "X‑HF‑Session‑Id" header lets the router keep a conversation on the same provider’s cache, improving response time for ongoing chats.
What the new routing does
OpenRouter measures every provider in real time and switches to another one if the first slows down or fails. According to OpenRouter’s own tests, the new routing cut capacity errors by 46 %, shaved 16.9 % off the time to first token, and lifted overall throughput by 16.3 %. Hugging Face passes the provider’s price straight through to users, adding no markup. Pro plans get a $2 monthly credit, and Team or Enterprise plans receive $2 per seat.
Impact on open‑weight models
Open‑weight (open‑source) models now dominate the traffic on OpenRouter, accounting for about 71 % of tokens processed by the end of September 2026, up from 26.8 % in October 2025. Open models cost roughly a quarter of what closed models charge per million tokens – about $0.28 in versus $1.27 for input, and $1.62 out versus $6.36. The best open model, Kimi K3.1 Pro, scores 46 on the Artificial Analysis Intelligence Index, while the top proprietary model, Claude Fable 5.2, scores 65 but costs about seven times more per task.
Why it matters
Developers can keep their existing code but get faster, cheaper responses automatically, which reduces operating costs for apps that use chat models. Users of open‑source models benefit from lower prices while still accessing the same API, making AI services more affordable for small teams and hobbyists.