The router became the product: Hugging Face hands Inference Providers to OpenRouter, and open-weight models hit 71.7% of the tokens it serves

Hugging Face moved Inference Providers onto OpenRouter's router, so provider selection, retries and failover now happen on someone else's infrastructure. Open-weight models went from 26.8% to 71.7% of the tokens OpenRouter served in a year, while the best open model still scores 46 against 65 on the Artificial Analysis Intelligence Index.

The router became the product

Hugging Face has handed its inference router to OpenRouter, and that is a larger change than a partnership announcement suggests. Every language model call through Hugging Face Inference Providers now lands on a router that OpenRouter built, and that router picks the provider for you, prices the call for you, retries the call for you, and steers your traffic away from a provider that starts failing. Your code does not change. The decision layer underneath it does.

The announcement landed on the Hugging Face blog on November 16, 2026 (Hugging Face and OpenRouter join forces). OpenRouter’s argument for the switch is scale: it already routes across more than 80 providers and over 600 models, serves more than 800 trillion tokens a month to over 10 million users, and therefore knows which provider is fast, cheap and reliable from live measurement rather than from a price table that goes stale overnight.

What changed underneath your code

Hugging Face switched Inference Providers from its own custom router to OpenRouter’s. A request now goes to OpenRouter, which picks the provider per call based on live price, speed and reliability, and retries on another provider if the first one fails. The API surface stays where it was: the router speaks OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages API for every chat model regardless of which provider sits behind it. Hugging Face ships setup guides for Claude Code, Codex, OpenCode, Pi and Hermes Agent, and the same Hugging Face token covers Storage Buckets, Jobs and dedicated Inference Endpoints alongside inference.

Three details decide how much this matters for a given codebase:

  • Provider choice is now a suffix on the model id. zai-org/GLM-5.3 is the default blend of price, speed and reliability; :fastest asks for highest throughput, :cheapest for lowest price, and :novita pins the request to a single named provider. Routing is also cache-aware, so a conversation stays on the provider holding its prompt cache; send a stable UUID per conversation in the X-HF-Session-Id header to keep that stickiness explicit.
  • Billing did not change. You pay the provider’s own price on your Hugging Face bill, with no markup added by Hugging Face. PRO accounts get $2 of monthly credits and Team and Enterprise organizations get $2 per seat.
  • Only language models moved. Image, video and audio generation still run directly through Hugging Face Inference Providers, without OpenRouter in the path, and the LoRA catalogue is untouched.

The open-weight share, and what it costs

The number that makes this consequential is not the partnership but the traffic mix behind it. On OpenRouter, open-weight models went 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 (OpenRouter, October 2026). Over the same period weekly open-weight volume grew more than sixty-fold, against almost tenfold for closed models. OpenRouter’s own analysis page frames the same shift slightly differently, as 27% to 72% of the tokens sent to the top 50 models, which is the more flattering denominator for the same underlying traffic.

The cost curve explains why the share moved. Open-weight models run about a quarter of the price on average: $0.28 per million input tokens and $1.62 per million output tokens, against $1.27 and $6.36 for closed models. The most-used model on OpenRouter this week is an open-weight one, DeepSeek V4.1 Flash 1017, with 59.2 trillion tokens served (rankings).

The benchmark gap that did not close

Cheaper and more popular is not the same as better. On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro, the best open-weight model, scores 46 against 65 for Claude Fable 5.2, at roughly a seventh of the cost per task ($3.68 against $26.54). The gap is real, it is measurable, and it did not close over the year in which open-weight traffic grew sixty-fold. OpenRouter’s own framing is that the best closed-weight models still score meaningfully higher, and that for a lot of workloads the trade is worth taking anyway.

That is the trade a router now makes on your behalf. Whether it makes the right one depends on the workload, and you can no longer see or adjust the decision unless you pin a provider yourself.

What OpenRouter rebuilt, and what it charged for it

The routing underneath is new, and OpenRouter published the method along with the result (How OpenRouter rebuilt provider routing). The old router mostly rewarded providers for one thing: a low price. The new one tracks each endpoint’s realized price, latency, throughput and error rate, and scores the blend of what a good answer actually costs: the price plus the waiting. Traffic moves off an endpoint that starts failing and returns once it recovers, and a failed attempt is retried on the next best provider inside the same call. Around 77% of requests on OpenRouter go to models served by more than one provider, so this is the common case rather than an edge case.

OpenRouter’s own measurements put the gains at a 46% reduction in capacity errors, a 16.9% improvement in time to first token and a 16.3% improvement in throughput. The same announcement carries the cost of that improvement: a 6.0% increase in cost. Treat all five figures as vendor-reported, measured on OpenRouter’s own traffic without a methodology a reader could independently re-run.

The breaking change that is already live

Hugging Face flags a breaking change: custom provider keys now behave differently, because OpenRouter performs the routing rather than Hugging Face’s own layer. Code that pinned a provider, or that assumed a specific vendor’s error semantics and latency profile, can behave differently after the switchover.

The tooling for seeing what happened is new as well. Model pages carry two tabs that did not exist before: Providers lists every provider OpenRouter can route the model to, with its input and output price, speed and uptime over the last 24 hours, and Metrics shows the best price, throughput, time to first token, context length and uptime among them. The usage overview breaks spend down by the provider each request actually went to. And in settings you can turn off any provider you do not want, at the account or organization level, and the router never sends requests there.

What to do about it

If you run inference through Hugging Face today, three concrete checks pay for themselves this week. First, run through your code and find every place that pinned a provider or relied on a specific vendor’s error and retry behaviour, because those are the call sites the breaking change lands on. Second, pin a provider explicitly for the workloads where latency is user-visible, using the suffix syntax, and let the default router handle the batch and background work where a 6% cost difference is cheaper than the time you spend tuning. Third, turn on the Providers and Metrics tabs for the models you depend on and read the 24-hour uptime numbers before you trust the default choice with production traffic. The routing decision is no longer visible by default, and the default is now the decision.

Sources