Is the Hugging Face inference API free in 2026? We counted.

The count

Those four numbers are the honest answer, and the second pair is where people go wrong. Not flagged free is not the same as costing money, and no price published is not the same as free. Three different things are being counted:

Who is actually behind it

The router is a front door for fourteen inference companies. Hugging Face runs almost none of this itself, which is the thing worth understanding: you are choosing a provider whether you realise it or not, and the same model has different prices, different context limits and different speeds depending on who serves it.

Every one of the 327 entries reported status: live when we counted, which is more than most catalogues can say. The routing metadata is genuinely good: per provider context length, tool support, structured output support, time to first token and throughput from a recent probe. If you are picking a provider for a model, that data is better than anything you will assemble yourself.

Calling it

It is OpenAI-compatible, so the only unusual part is the base URL and the model id, which carries the provider after a colon when you want to pin one.

curl https://router.huggingface.co/v1/chat/completions \
  -H "Authorization: Bearer $HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Llama-3.3-70B-Instruct",
    "messages": [{"role": "user", "content": "hello"}]
  }'

Get the token from huggingface.co/settings/tokens. Create a fine-grained token and enable Make calls to Inference Providers; a classic read token works too but grants more than this needs.

And the command behind every number on this page, which needs no token at all:

# how many entries are flagged free?
curl -s https://router.huggingface.co/v1/models \
  | jq '[.data[].providers[] | select(.is_free == true)] | length'

So what should you use?

If you want a specific open model on a specific provider and you are willing to pay for it, the router is a good piece of engineering and the comparison data is the best part of it.

If what you actually wanted was the old thing, a free endpoint you can point a script at, that is the part that went away. We run one: DreamPrompting is a free OpenAI-compatible gateway that routes across about two dozen providers' free tiers, including this router, and moves to the next provider when one is busy rather than failing.

curl https://dreamprompting.com/api/v1/chat/completions \
  -H "Authorization: Bearer $DP_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"hello"}]}'

Related, and built the same way, by calling things rather than trusting their docs: Cloudflare Workers AI lists 31 text models and we could call 28, and why your rate limit is measured in tokens rather than requests.