Hugging Face Routed Everything Through OpenRouter. Free Users Get Zero

A better router, a wider catalogue and the same token. Also a quietly closed door for anyone who was reaching open-weight models without paying.

Share
Hugging Face Routed Everything Through OpenRouter. Free Users Get Zero

Hugging Face moved Inference Providers onto OpenRouter's router on 16 November 2026. Every open-weight model a provider serves is now callable with one Hugging Face token, routed across 80+ providers on live price, speed and reliability, with automatic retry inside the same call. Prices are unchanged and carry no Hugging Face markup. What did change: free accounts now receive no monthly inference credits, and custom provider keys stopped working for chat models.1

Both of those last two facts are in the announcement, stated plainly and without emphasis. Together they close the last route by which someone with no budget could run open-weight models through Hugging Face on their own terms.2

What the new router actually does

ChangeDetail
RouterOpenRouter replaces Hugging Face's custom router for chat models
Provider pool80+ providers, selected per request on price, speed and reliability
Failure handlingRetries on the next best provider inside the same call
API compatibilityOpenAI Chat Completions and Responses, Anthropic Messages
Routing controlSuffix the model id with :cheapest, :fastest or a provider name
PriceProvider list price, no Hugging Face markup
Custom provider keysNo longer accepted for chat models
Image, video, audioStill direct through Hugging Face, not routed via OpenRouter

Prices measured on 27 October 2026. Vendor plan terms, allowances and prices change without notice, so treat every figure on this page as a dated snapshot rather than a permanent fact. Where a number here differs from the vendor's own pricing page, the vendor's page is correct and the figure above is simply older than it.

The engineering is straightforwardly good. OpenRouter rebuilt its routing to weigh a provider's live performance rather than just its list price, and reports that this cut capacity errors by 46% while improving time to first token by 16.9% and throughput by 16.3%.

Routing is also cache-aware, so a conversation stays with the provider holding its prompt cache. A stable UUID per conversation in an X-HF-Session-Id header makes that explicit, which matters because moving a cached conversation between providers silently multiplies cost.

The credit table, read carefully

Here is the line that reorganises everything. Hugging Face's pricing documentation lists monthly included credits by account type, and the row for free users reads: none, with extra usage described as requiring a credit purchase.

AccountMonthly creditsCan be spent on
FreeNoneNothing, a purchase is required
PRO$2.00All Hugging Face compute services
Team or Enterprise$2.00 per seatAll Hugging Face compute services

Two dollars is a real allowance and it is honest money, because the same credits also pay for Inference Endpoints, upgraded Spaces hardware including ZeroGPU beyond the quota, and Jobs. It is not, however, very much inference. At the open-weight averages in Hugging Face's own analysis, roughly $0.28 per million input tokens and $1.62 per million output, two dollars is around seven million input tokens or a little over a million output tokens a month. Enough to evaluate, not enough to run anything.

Someone on a free account gets none of that. In our comparison of free inference tiers, this is the cleanest example of a platform where the free plan is now a browsing and hosting plan rather than an inference plan. The ZeroGPU quota on Spaces still stands at five minutes a day, which remains the only meaningful free compute on the platform.

Bringing your own key no longer works

This is the part with no workaround. Because chat requests now route through OpenRouter and bill to your Hugging Face account, you cannot attach your own API key for a provider on these models. Nominally the provider suffixes and the per-provider toggles still work, so you can steer where a request goes, but you cannot pay that provider directly.

Anyone who was using Hugging Face as a convenient front door to a provider account they already funded has lost that arrangement for chat models. Custom keys continue to work for image and video providers, and a comment in the announcement says connecting an OpenRouter account to Hugging Face is on the roadmap, which would restore a version of the old option. It is not here yet.

Why did Hugging Face make this switch?

The strategic logic is not mysterious. On OpenRouter, open-weight models went from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026, peaking at 77.5% in July. Weekly open-weight volume grew more than sixtyfold over that period, against roughly tenfold for closed models.

The price gap is the reason. According to Hugging Face's own analysis, the averages are $0.28 per million input and $1.62 per million output for open-weight, against $1.27 and $6.36 for closed models, roughly a fourfold difference across the board.

The quality gap is real but narrowing at the top end. On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro scores 46 against 65 for Claude Fable 5.2, at about a seventh of the cost per task, $3.68 against $26.54.

Who should care

If you already pay for PRO or sit in a Team organization, this is an upgrade with no downside. The same $2 per month now buys routing across 80+ providers with automatic failover, and a wider catalogue than the old custom router exposed.

If you are on a free account, the practical consequences are these: chat inference now requires a credit purchase, the same position Together AI already took, and your own provider keys will not substitute. The verdict is that the routing is genuinely better and the access is genuinely worse, and the free plan was the part that moved. The limitations are the zero credit line and the closed custom-key path. In our reading the fix is either a small credit purchase or a move to a platform that still exposes a free allowance.

Frequently asked questions

Do free Hugging Face accounts get inference credits?
No. The pricing table lists none for free users, and credits must be purchased.

How much monthly credit does PRO include?
$2.00, usable across Inference Providers, Inference Endpoints, Spaces hardware and Jobs.

Can I use my own provider API key?
Not for chat models any more. Custom keys still work for image and video providers.

Did prices go up?
No. You pay the provider's list price and Hugging Face adds no markup.

Does my code need to change?
No. The same token and the same APIs work, including OpenAI and Anthropic compatible formats.

Are image and video models routed through OpenRouter?
No. Image, video and audio generation still run directly through Hugging Face Inference Providers.

The router switch, the retry behaviour, the routing suffixes, cache-aware routing, the compatibility statement, the pricing table, the no-markup statement, the custom-key breaking change and all market statistics are from Hugging Face's announcement of 16 November 2026, whose source is also published on GitHub. The included-credit table and what the credits can be spent on are from Hugging Face's pricing and billing documentation, read on 17 November 2026.

Sources

  1. Hugging Face Blog: Hugging Face and OpenRouter: huggingface.co/blog/openrouter. Used for: Router switch to OpenRouter on 16 November 2026; 80+ providers; retry inside the call; :cheapest/:fastest/provider suffixes.
  2. Hugging Face Docs: Inference Providers pricing: huggingface.co/docs/inference-providers/pricing. Used for: Free users get no monthly credits; PRO and Team/Enterprise get $2.00 per seat; custom provider keys closed for chat models.

All sources are the vendor's own plan or documentation pages, checked on 27 October 2026. Plan terms change without notice, so confirm current limits before you rely on them.