Shared model APIs

Not open yet. This is how it will work. Numbers and rules may still change before it opens.

Using a shared model

The model name is api/<name>. It is listed apart from our own models.
You agree first. These requests are handled by someone else's account and their upstream service, and they can read your prompt. You can agree for the whole account, or for one request only.
Public content only. Requests at a higher level are refused, not sent on. Free trial accounts cannot use shared models.
Charged at the model's price. If a provider's reported usage does not match the length of the text, we bill an estimate instead, so you are never charged for tokens that were not there.
When no provider is available, you get "no such model" or "no provider can answer right now", and nothing is charged.

You need a full account, an upstream service that speaks the OpenAI API, and your own upstream key.
Run the provider program on your own computer. It connects out to Babel Mesh, so no port needs to be opened. Stop it when you do not want to share.
Your upstream key stays on your computer. The program uses it only for requests to your upstream. It is not logged, and it is never sent to Babel Mesh. Requests from us arrive without the requester's identity.
Set a daily limit. When it is reached, the program stops taking work and starts again the next day.
State the model's tier honestly. Tiers are checked, and an inflated claim is found.

A new model starts on trial. During the trial it only receives check questions, generated by a program and graded automatically.
Real requests come only after it passes, and check questions keep being mixed in from time to time.
A model caught pretending is isolated. It gets no more real requests, and earnings still in their hold are forfeited.
Some pretending is hard to spot. If an upstream silently swaps a model for another of the same tier, we cannot guarantee we will tell.

The provider keeps part of what is paid, 60% by default. This can be adjusted.
Earnings are held for 72 hours before they become available. If the model is isolated during the hold, those earnings are forfeited.
Available earnings go to your account's credits.

Providers and their upstream can read what you send. Do not send passwords, ID numbers, trade secrets or private personal data through a shared model.
Judge the upstream provider's terms yourself. Many providers do not allow reselling or sharing API credit. The consequences are yours, and Babel Mesh does not decide this for you.
We check. A small part of the requests are run again on our own model and the results compared, to catch pretending.