Article

Move Hugging Face language models off provider-key billing

For teams using Hugging Face language models, check who pays, how credits are shared and when provider choices disable automatic failover.

Editorial illustration for Move Hugging Face language models off provider-key billing: a geometric block represents a model release. Not documentary evidence.

Hugging Face announced on November 16, 2026 that custom provider keys no longer work for language models accessed through Inference Providers. Those requests now bill a Hugging Face account or organization, with OpenRouter selecting the serving provider. Affected teams should check their Hugging Face token, payer and funding before relying on the old setup.

Which requests need attention?

Start with language-model requests made through Hugging Face, not calls sent directly to a provider’s own API. Custom keys for image and video providers remain supported. Audio generation also keeps its existing routing, but the announcement does not explicitly include audio in the custom-key exception.

The announcement describes the change as already in effect; it gives no future migration deadline. Connecting an existing OpenRouter account is listed as a future feature, not an available way to carry an OpenRouter balance into Hugging Face.

Keep your Hugging Face token and choose who pays

OpenRouter says existing Hugging Face tokens and code keep working. You do not need a separate OpenRouter credential for the documented setup. That compatibility does not preserve the removed provider-key payment route.

Check that your Hugging Face token has the Make calls to Inference Providers permission. The current HTTP quick start still uses https://router.huggingface.co/v1/chat/completions. The announcement does not prescribe a blanket token rotation or a minimum client-library version.

For Team or Enterprise billing, members still authenticate with their own user tokens, but must explicitly select the organization:

  • HTTP requests: set X-HF-Bill-To to the organization name.

  • Hugging Face Python client: set bill_to when initializing the client.

  • Hugging Face JavaScript client: set billTo in the client options.

Selecting the payer also selects a provider policy: organization-billed requests follow the organization’s settings instead of the member’s personal settings. A successful personal request therefore does not prove the same provider is permitted for a team-billed request.

Check available credits and the payment method

The credit table checked on November 16 lists no monthly allowance for Free accounts, US$2 for PRO, and $2 per Team or Enterprise seat. The organization allowance is shared. These are compute credits, also usable for Endpoints, Spaces and Jobs—not a reserved language-model budget.

Illustrative calculation: an existing five-seat paid organization receives 5 × $2 = $10. Assume no purchased balance remains and other eligible compute has used $7, leaving $3. If subsequent inference costs $8, another $5 needs funding. This is arithmetic, not measured usage or a model-price quote; subscription fees, tax and negotiated arrangements are excluded.

Check the payer’s billing dashboard and credit expiry dates before adding funds. Monthly included credits apply before chargeable usage. Organizations already using active cloud-marketplace billing cannot buy prepaid credits; their metered usage goes to the linked cloud provider instead.

For marketplace approval and switch-date details, see the Hugging Face Google Cloud billing guide. Its earlier general description of provider-key billing no longer applies to the language-model requests covered here.

Team and Enterprise administrators can set an Inference Providers spending limit. Review it before moving traffic; the documentation does not specify enforcement timing or possible overshoot.

Choose a provider without accidentally losing failover

Provider selection and provider payment are now separate choices. Pinning a provider does not restore billing through its custom key.

The routing documentation gives four choices:

  • No suffix: automatic selection balances price, speed and reliability.

  • :cheapest: prioritize the lowest price.

  • :fastest: prioritize output throughput, measured in tokens per second—not necessarily the shortest wait for the first token.

  • A provider suffix: restrict the request to that provider. A disabled provider, or one that does not serve the model, produces an error.

Automatic routing can retry another provider when one fails. A pinned provider is never replaced. If several providers are acceptable, use the enabled-provider settings to define that set without fixing every request to one destination. If one specific provider is required, plan for errors instead of cross-provider recovery.

Check each permitted provider’s data-handling policies as well as its model support. Central billing does not establish identical retention or processing-location terms.

Hugging Face says it passes through provider prices without markup. That is not a promise of an unchanged total bill: default routing can choose a more expensive provider for speed or reliability, and spending also depends on request volume and cache reuse.

Check usage after the billing change

For repeated conversations, Hugging Face documents automatic prompt caching—reuse of processing for unchanged text. Its Python and JavaScript clients handle caching automatically. For raw HTTP, send a stable UUID per conversation in X-HF-Session-Id to keep requests on the same provider; other value formats are ignored. Check the response’s usage.prompt_tokens_details.cached_tokens rather than assuming every turn reused a cache.

Teams also using AWS can consult the separate Bedrock prompt-caching cost guide for its write/read charging model. Those prices and controls do not apply to Hugging Face.

Before scaling traffic, confirm the intended payer, allowed providers and funding, then inspect the usage breakdown by model and provider after a small representative workload. A returned answer establishes that a request worked; it does not by itself establish that the right account paid or that costs stayed within your plan.