How the built-in prompt works

Every model has its own prompt that shapes its replies. You buy it once, and it works with all your keys for that model.

What happens to a request

1. You send a regular request with your key.
2. Our server puts the model's prompt first.
3. Your system message (if any) follows, then your messages.
4. The model answers with both in mind.

The prompt is stored encrypted on the server and is never sent to you or included in replies. If the model tries to repeat it word for word, that reply is replaced.

The key's model is fixed

A key is issued for one model. A request with a different model value returns 403 model_not_allowed.

Updates

When a model's prompt is improved, the new version starts working on all existing keys right away. You don't need to change anything.

Prompt tokens

The prompt is part of the model's input, so its tokens count toward the request's input tokens and are billed at the input price. See Pricing and billing.

The prompt as a .txt file

If you need the prompt text itself to use elsewhere, the shop sells it separately: a .txt file with the prompt and a guide. Each file carries a personal buyer mark, so publishing or reselling it is not allowed.