SateiStart saving

Free test

Is that endpoint really what it says it is?

Hand us an OpenAI-compatible endpoint and we ask it the same six questions every listed host faces, in the same shifting wording. Back comes its answers, how tightly they hold, and the tokens it claimed.

Used once, to make the check calls, and never stored.

What you get back

  • The answers. Six questions, retold in shifting wording, so the endpoint's favourites surface.
  • A stability score. How tightly the endpoint repeats itself across the whole run.
  • Failures and refusals. How many calls never came back clean.
  • A verdict. Holds up, too close to call, or off pattern, with the rows behind it.

Nothing is published from a test you run. Want a model measured across every listed host instead? See the price ledger.

How it runs

Three fields, about half a minute

01

Point it at a base URL

Any endpoint speaking the OpenAI chat format: your own box, a host's direct API, a proxy.

02

Name the model

Use the id your endpoint expects. It faces the same six questions as every listed host.

03

Read the verdict

Holds up means it behaves like one steady model. Off pattern means its own answers disagree.