Decision models that learn from experience. Built to decide. Made to learn.

An octopus at an old telephone switchboard, a receiver at its ear, plugging calls through with its other arms.
#1decision model on JevBench capability
160 msmedian per decision, end to end
$0.02per million tokens, kettei core

Call the kettei models. Fine‑tune them on your data.

Our models

A model for each decision, and it never stops learning.

01

Sized to the decision

Four sizes, from fat for the hardest calls to dot. Use the smallest one that gets it right.

02

Trained on your data

Make it yours. Fine-tune on your own data in minutes.

03

Keeps learning in production

Continual learning. Every outcome in production feeds the next round of training.

Use cases

The right call, right away.

Customer support · escalationketteiGPT‑6 Luna

More accurate82.5%72.7%
4× cheaper$0.008$0.032
17× faster163 ms2.8 s

Safety · prompt injectionketteiLlama Prompt Guard 2

12× safer61%5%
Same price=
Zero false alarms100 / 100let through
An octopus on a stool in a railway signal box, drawn in ink: one arm pulls a lever in a row of levers while others hold a mug and write in a notebook. Only that lever, and the signal it sets outside by the track, are in colour.

Continual learning you can trust.

Each new model joins the family only if it passes its gate.

Why System One

Most steps need a decision, not an essay.

SYSTEM ONE · FAST AND CHEAPSYSTEM TWO · SLOW AND EXPENSIVE