Platform

How Kael works

What happens inside Kael between your request and the answer, and when to choose it.

From request to answer

You make one API call. Inside Kael, a request is handled in up to two passes before anything comes back.

1. Draft

   A first answer. Kael also rates the request easy or hard.

2. Review

   The draft is checked against your request. Skipped for easy requests.

3. Answer

   Only the reviewed result is returned to you.

The review pass reads the draft against what you asked and confirms it, corrects it, or replaces it. For code, that is where bugs and security problems the first pass missed get caught. The number of internal calls can change as Kael improves. Neither your bill nor the usage you see depends on it.

What that means for you

• It is slower than a single call, on purpose. Draft text is never shown, so the first answer token comes after the draft finishes. A stream sends short thinking titles in the meantime. For the fastest replies, set thinking to off.
• Easy requests are faster. When a request is rated easy, the review is skipped. The price per token does not change.
• A failed review does not fail your request. If the review pass errors or times out, you get the draft instead.
• You are billed once. Your input and the final answer, nothing else. See Pricing and billing.
• Output is not deterministic. Sampling settings are ignored, so the same request can return different answers.

What Kael is built on

Kael is a Composite Intelligence System: several AI components coordinated behind one API, instead of one end-to-end model. It is built on open-weight models and is not trained from scratch. Quancis does not publish which models, and Kael will not say. Kael's own weights are not available to download.

Safety screening

Requests and answers can be screened against the acceptable-use section of the Terms of Service. A flagged request can be held for review and the account paused while it is checked. See Data and privacy.

When to use Kael