Hugging Face Inference API Pricing Calculator

1. Pick the billing model: Inference Providers per token, HF Inference per compute second, or a dedicated Inference Endpoint per hour.
2. Enter the prices and your usage. Token prices come from the provider row for your model. Endpoint rates fill in when you pick an instance, and every price stays editable.
3. Choose your account type so the monthly credits are taken off, then read the cost and the endpoint comparison.

What each result tells you:

The subscription fee for a PRO, Team or Enterprise plan is a separate charge. It is not part of the result.

Hugging Face inference pricing is the cost of running a model through Hugging Face, and today it comes from two products with different billing. Inference Providers is the serverless, pay-as-you-go service. Inference Endpoints is the dedicated service billed by the hour.

The name in this page's title is the older one. Hugging Face's billing guide says the HF Inference provider used to be called Inference API (serverless) before Inference Providers existed. HF Inference is now one provider among many inside Inference Providers.

Hugging Face states that the free-user credit is subject to change. Free-user credits can be spent on Inference Providers. Credits that come with a PRO, Team or Enterprise plan are general compute credits, which also cover Inference Endpoints. The PRO plan is listed at $9 a month.

Hugging Face says that, as of July 2025, HF Inference focuses mostly on CPU tasks such as embeddings, text ranking and text classification. Large language models are mostly served by partner providers at per-token rates.

The calculator applies one of three formulas, then subtracts your monthly credits.

1. Per token: tokens per request x requests per month gives monthly tokens. Each total is priced per 1 million tokens.
2. Compute time: seconds per request x the hardware price per second x requests per month.
3. Endpoint: the hourly rate x the instance hours billed. This is the formula in the Hugging Face endpoint pricing guide.
4. Monthly cost = usage cost - credits. The result stops at zero.
5. Break-even requests = the instance rate x 730 hours / your cost per request.

Token counts are your own figures. A fast way to turn text length into tokens is the AI token calculator.

A free-tier user sends 100,000 requests a month through Inference Providers. Each request has 1,000 input tokens and 300 output tokens. The provider row lists $0.20 input and $2.50 output per 1 million tokens. That row is one example read from the supported models table on 2026-10-02, not a typical rate.

Meaning: the bill is $94.90 for the month. A PRO user with $2.00 of credits would pay $93.00 plus the plan fee. A dedicated NVIDIA T4 endpoint left on all month costs $365.00, so pay per use stays cheaper until about 384,211 requests a month.

The Provider You Route To

The same model is priced differently by each provider. Hugging Face passes the provider rate through with no markup, so the row you pick in the supported models table sets your price.

Output Tokens

Output usually costs more than input. In the example, output is 30 million of 130 million tokens but $75.00 of the $95.00 bill.

Routing and Keys

Requests routed by Hugging Face are billed on your Hugging Face account, and monthly credits apply. With a custom provider key, the provider bills you directly and the credits do not apply.

Running Hours

A dedicated endpoint is billed for the time it runs, not for requests. An NVIDIA T4 instance on for 8 hours across 22 days costs $88.00 instead of $365.00.

Replicas and Autoscaling

Each replica is billed at the full rate. The pricing guide's example scales a T4 endpoint from 1 to 3 replicas for 15 minutes every hour. That adds 182.5 hours x 2 replicas and brings the month to $547.50.

Instance Size and Cloud

Rates differ by hardware and by cloud. One NVIDIA A100 is $2.50 an hour on AWS and $3.60 an hour on GCP in the official table.

Inference Providers cost scales with requests, and an Inference Endpoint costs the same whether it serves one request or a million. The right choice depends on steady volume. The table shows official hourly rates and the cost of one replica left on for 730 hours.

The break-even point is the endpoint's monthly cost divided by your pay-per-use cost per request. At $0.00095 a request, the $365.00 T4 endpoint matches pay per use at about 384,211 requests a month. The same sum for any fixed cost is covered by the break-even calculator.

What Does HF Inference Cost per Request?

HF Inference bills compute time. The billing guide's example is a request that runs for 10 seconds on a GPU machine costing $0.00012 a second, which is billed $0.0012. Fifty thousand such requests cost $60.00.

What Should You Budget?

Budget from measured usage. Run a short test, read the token counts or request times, and enter the averages. One instance has a limit on how many requests it can serve, so a busy service needs more replicas than the break-even sum suggests.

Choosing Between Serverless and Dedicated

A team with steady traffic compares its pay-per-use bill with an always-on endpoint. The break-even result shows which side of the line it is on.

Comparing Providers for One Model

A developer enters two provider rows in turn. The difference in monthly cost shows what a faster provider costs. Response speed matters as well as price, and the token generation speed simulator shows how long a reply takes at a given speed.

Sizing a Side Project

A hobby user checks whether the monthly credits cover the project. A result of zero means no pay-as-you-go charge at that volume.

Planning Working-Hours Endpoints

A team that needs a private model only during the working day enters the real running hours. The saving against an always-on endpoint is the difference.

1. Expecting One Price per Model

There is no single Hugging Face price for a model. Each provider sets its own rate, and the rates change.

2. Treating the Free Credit as a Free Tier for Production

The free-user credit is $0.10 a month and is subject to change. It suits testing, not a live service.

3. Leaving an Endpoint Running

A dedicated endpoint bills for every hour it runs, with or without traffic. An idle T4 instance still costs $365.00 a month.

4. Forgetting Extra Replicas

Autoscaling adds replicas at the full hourly rate. Count the scale-up hours, not only the minimum replica.

5. Using a Custom Key and Expecting Credits

Credits apply only to requests billed by Hugging Face. A custom provider key moves the bill to the provider.

6. Reusing Old Price Tables

Hugging Face has renamed and restructured these products. Older per-model price lists for the serverless Inference API do not describe billing today.

The arithmetic is exact for the prices and usage you enter. The estimate depends on those inputs, and provider rates change without notice.

Only up to the monthly credits. Free users get $0.10 of Inference Providers credits a month, subject to change. After that, usage is pay-as-you-go and needs purchased credits.

The serverless service is now Inference Providers. Hugging Face says its own HF Inference provider used to be called Inference API (serverless). Dedicated deployments are a separate product named Inference Endpoints.

No. Hugging Face states that it charges the same rates as the provider and passes the costs through with no extra fees.

It is the hourly rate x the hours it runs x the replicas. One AWS NVIDIA T4 instance at $0.50 an hour costs $365.00 for 730 hours, and the smallest listed AWS CPU instance at $0.033 an hour costs $24.09.

A PRO account includes $2.00 of compute credits each month, and the plan is listed at $9 a month. The credits can be spent on Inference Providers, Inference Endpoints and other Hugging Face compute services.

A dedicated endpoint is cheaper once your monthly pay-per-use bill is higher than the endpoint cost. At $0.00095 a request, that point is about 384,211 requests a month for a $365.00 endpoint.

Prices are shown by the hour, and Hugging Face states that the actual cost is calculated by the minute. This tool works in hours, so enter part hours as decimals.

No. With a custom provider key the provider bills you directly, and Hugging Face monthly credits do not apply to those requests.

Hugging Face publishes a supported models table for Inference Providers. It lists the input and output price per 1 million tokens for each model and provider.

It does not suit image, video or speech workloads priced per image or per minute by a provider. It also does not size the hardware a model needs.

No. The calculation runs in your browser and nothing is sent to our servers. Anything you Save stays in this browser only.

• Inference Providers: pricing and billing (Hugging Face documentation, checked 2026-10-02).
• Inference Endpoints: pricing (Hugging Face documentation, hourly instance rates and cost formula).
• Hugging Face pricing (Hugging Face, plans and hardware prices).
• Inference Providers: supported models (Hugging Face, per-model and per-provider token prices).