OpenAI introduced its Decisions API at its DevDay conference on 29 September, The New Stack reported. Built on the Luna model, the limited-preview service takes questions, possible answers and context from developers, then returns a choice with a confidence score rather than generating a written reply.
Key points
- Developers supply the questions, possible answers and context for each decision.
- OpenAI says a result takes 150 milliseconds, against 1.6 seconds for a comparable GPT-6 Luna request.
- The service is in limited preview, with a broader release planned after 29 September.
Luna chooses from developer-supplied answers
The Decisions API uses Luna, the smallest and least expensive model in OpenAI’s current lineup, according to The New Stack. The Decoder described the service as using a version of GPT-6 Luna. OpenAI had released GPT-6 Sol and Luna with lower API prices and fewer errors, as AI Affairs previously reported; this API gives Luna a narrower job than composing an answer to an open-ended request.
That job begins with a list of permitted answers. A developer supplies the question, the choices and the context, and the service returns scores for those choices. It is closer to sorting mail into labelled trays than writing a letter: the destinations are set before the request arrives. The New Stack identified classifying content, routing requests and selecting an agent’s next action from a restricted menu as potential uses.
Sorting incoming messages could become a choice among permitted categories rather than a request for a written explanation. The score beside an answer could help determine whether a classification warrants another look, though the usefulness of that distinction would depend on how well the scores correspond to correct answers.
OpenAI says the decision service returns a result in 150 milliseconds, compared with 1.6 seconds for GPT-6 Luna, The New Stack reported. Those are OpenAI’s timings, rather than measurements of a developer’s application under load. The comparison concerns the time taken to obtain an answer, a practical consideration when a choice must be made before another part of an application can proceed.
TypeSafe’s Jev puts scores in focus
OpenAI’s announcement follows the mid-September introduction of Jev, a decision model from TypeSafe AI, The Decoder reported. The New Stack described the Decisions API as a likely response to Jev. That is an account of the competitive timing, rather than a reason OpenAI has given for building the service.
A conventional chat model can also be prompted to choose from a list. The difficulty, as The New Stack described it, is that a confidence score attached to its answer may be a rough guess, while producing the answer can consume more tokens. A small classifier offers another route, but it needs labelled training examples and must be retrained when its categories change. A decision model accepts the permitted answers with the request and returns scores without that retraining step.
The distinction matters when the available choices change. A fixed classifier can sort reliably among the categories it was trained to recognise, while a prompt can introduce a new list without another training run. The Decisions API is intended to combine the latter flexibility with an output that software can use as a scored choice. The New Stack said decision models can provide more realistic confidence scores than chat models; how useful any score is still depends on the particular decisions it accompanies.
OpenAI begins with a limited preview
The limited preview gives developers a defined interface for asking a model to pick from supplied options, but several details matter to anyone considering it as a routine part of an application. The New Stack said the price per call remains unclear, as do the number of candidate answers a request can contain and whether developers can tune the service on their own data. Each bears on a different decision: cost at volume, the size of the menu and whether the scores can be adapted to a particular task.
OpenAI already offers another scored output through its Moderation API. That service returns scores by category, but OpenAI sets the categories. With the Decisions API, developers provide the answers themselves, making the choice of categories part of the request rather than a fixed part of the service.
OpenAI plans a broad release in the days after 29 September, The New Stack reported.
