OpenAI has made Decisions API available to all developers in public beta after a limited preview. Powered by GPT-6 Luna, the endpoint returns structured decisions that applications can use to route requests, classify content, evaluate inputs or select an agent’s next step. Developers can supply context as text or images.
Three formats for rules, choices and evaluation
Developers define the required decision format in each request: a predicate, a set of allowed options or a numeric scoring range. They then attach the context the model should evaluate. The structured response lets applications handle conditional decisions, multiclass classification and numeric evaluations without parsing generated prose or repairing malformed JSON.
Predicate returns an estimated probability that a specified statement is true, rather than simply a yes-or-no answer. The documentation lists policy checks, conditional gates and risk detection as uses for this format.
Choice selects from options defined in advance by the developer and includes confidence scores. It is intended for classification, request routing and action selection. Score instead produces an evaluation within a specified numeric range, supporting ranking, quality assessment and grading against a rubric.
From request routing to working with screenshots
Preview users also applied the API to visual tasks, comparing images and choosing useful frames from video. They used screenshots to select interface controls or form actions.
Decisions is intended for tasks whose answers can be expressed as a predicate, a finite set of options or a numeric range. Responses API remains available for generating prose and open-ended reasoning over long context.
OpenAI’s demo compared speed with Responses API
In its launch demonstration, OpenAI routed 10,000 customer requests into Billing, Technical and Sales categories. The company reported 150 milliseconds per request through Decisions API, compared with 1.6 seconds through Responses API—a roughly 10.7-fold speedup in that demonstration.
Every’s comparison with Jev produced mixed results
Every, which had preview access to the endpoint, compared Decisions with Jev. During an offline replay of computer-use tasks, Every reported that Decisions selected the correct control in 76 of 78 steps. Jev selected correctly in 73 steps in the same comparison.
In a separate thread-classification test, Every reported equal accuracy for both systems, with Jev returning results faster. Every also said Jev performed better across its broader evaluation set.
Application integration and confidence scores
For existing OpenAI customers, Decisions integrates with the infrastructure they already use: it shares the SDK, authentication and billing infrastructure of the company’s other APIs.
Confidence values are model estimates. AlphaSignal notes that production decision thresholds need validation against representative labeled data, taking the consequences of different error types into account.
