Decision models
Found this helpful? Share it:
Found this helpful? Share it:
A decision model does not write text. You send it a state (the content to judge) and one or more named, closed questions, and it answers with typed values: a probability, a label, a score, or a bounding box. There is nothing to read and nothing to parse out of prose, which is the point: a gate that has to be read is a gate that can be argued with.
noul returns yes or no as a single calibrated probability. choice returns one
label out of a closed set, with a probability per label. score returns a number
on a rubric you supply, for thresholding. box returns a bounding box, for
locating the thing you asked about in an image. Not every model answers every
type; the model page lists what each one supports.
Pricing follows from the shape. The model bills per input token only: there are no output tokens to charge for, because the answer is a handful of numbers. The model directory prints that single price per model.
The native endpoint is POST /api/v1/decisions. The body is a state plus a
record of named questions, and the response carries typed answers and usage. The
same class is reachable through the OpenRouter-compatible door, so an SDK pointed
at the OpenRouter base URL can call it without a special case.
Routing, privacy floors and bring-your-own-key apply to a decision call exactly as they do to every other modality: the request goes to the provider your workspace is allowed to reach, and a floor that excludes a provider excludes it here too. A decision model that cannot be served under your floor is not offered rather than silently substituted.
Decision models are captured and compared in the evals pipeline like every other modality. A capture records which question was asked, the answers that came back, and the token cost, so two models can be compared on the same probe rather than on a vendor's claim. The directory page carries a static preview of the request shape; it makes no model call, so you can see the wire format before spending anything.
Was this helpful?