Listing your model as a provider
An inference company can sell its endpoint on llm.buzz at its own price. The steps from sign-up to the first paid request: the application, the listing, the automated check and the monthly statement.
Every number on this page comes from llm.buzz’s public catalog, read 27 October 2026, 05:52 UTC.
On llm.buzz a model is served by one or more providers, and any company that serves models on a public HTTPS endpoint can become one. Right now 11 providers serve the catalog, 1 of them independent. A listing that goes live is one more provider of a public model, for every caller, at your prices.
What it costs you
llm.buzz keeps none of your listed price: you earn it on every token you serve. Each month you get a statement of what you are owed, at the prices recorded on each request when it was admitted.
Sign up
Sign up, or sign in, on the provider sign-up page, and create a supplier organization: the console calls a provider a supplier. You become its owner. A supplier organization has no keys or credit of its own, and its owners and admins invite the rest of your team, as in any organization.
Apply
The Application page takes, once:
- the name callers see, your company’s legal name, country, postal address, an optional tax or VAT number, and a contact;
- your company’s logo: a PNG, WebP, JPEG or SVG file of at most 256 KiB;
- an optional description that callers see under your name once it is approved: one paragraph of plain text, at most 300 characters, with no links;
- acceptance of the current supplier terms;
- 1 to 10 endpoints.
Each endpoint has a name, an API (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Google Gemini or AWS Bedrock), a base URL, its key, its residency (EU or not) and its data policy: whether you train on prompts and how many days you keep them. The data policy has no default, because callers and routing rely on it.
- An endpoint’s name is your provider name, alone or followed by a space and a suffix, such as
AcmeorAcme EU. - Its base URL is a public
httpsURL, such ashttps://api.example.com/v1. - Its key is sealed when it arrives. From then on the console shows only its last four characters.
A person reviews every application. A rejection comes with a note on your Application page; change the application and submit it again. Once approved, you add more endpoints without another review.
List a model
On the Models page, List a model takes one of your endpoints, the model and a declaration. The model is either one already on llm.buzz, which must be a text model (it outputs only text and takes only text and images), or a new name, which an operator approves once.
The declaration is the upstream model ID your endpoint is called with, the precision you serve (such as fp8 or bf16), the context window, the output ceiling, and five prices per million tokens: input, output, cached input, cache write and reasoning.
You can also list a speech-to-text or text-to-speech model already on llm.buzz, from an OpenAI Chat Completions or OpenAI Responses endpoint that serves /audio/transcriptions or /audio/speech beside its chat path. Its declaration has no token limits and one price: per million seconds of audio, or per million characters of input text.
Pass the check
Before a listing goes live, an automated check sends your endpoint real requests, built as a caller’s request to the model would be:
| Step | When | Passes when |
|---|---|---|
stream | Always | The stream ends with a stop or length finish and carries text. |
usage | Always | Usage reports input and output tokens, with input between half and twice llm.buzz’s own count. |
tools | The model declares tools | The answer calls the given function with valid arguments. |
structured_output | The model declares structured output | The text is a JSON object in the given schema. |
image_input | The model declares image input | The answer to a small image carries text. |
prompt_caching | Your cached-input price is below your input price | A repeated prefix reports cached input tokens. |
An audio listing’s check runs its own steps instead: a short tone transcribed as json, and as verbose_json when the model answers it, for speech to text; “Hello.” spoken as mp3 and as pcm, and at another speed when the model takes speed, for text to speech.
A passing check puts the listing live. A failing one shows the first failed step’s reason, and the check runs again later on its own. Checks are never billed.
While you are live
Each request goes to the best healthy provider of the model, ranked by the model’s routing preset, so your price and your speed both count. On a model with the llm.buzz provider, the smart preset tries llm.buzz first and the other providers, yours included, are its failover. If your endpoint fails a request, the next provider answers it. A listing whose success rate under real traffic falls too low is paused, and returns when a check passes again.
A price cut applies from the next request. A price rise waits a notice period, and callers see the coming prices and their date on the model page. A request keeps the prices in force when it was admitted, so a change never reprices a request already running.
Get paid
When a UTC month closes, you get its statement: one line per listing with its paid requests and their tokens. You download it as CSV or print it, upload your invoice for it, and are paid outside the console. Failed attempts are not paid.
Ready to start? Read Become a provider, or go straight to the sign-up page.