OpenAI: GPT-6 Luna Decisions

GPT-6 Luna in OpenAI’s Decisions mode evaluates supplied evidence and returns typed answers about it. The producer documents GPT-6 Luna as the model for this specialized decision interface. It can estimate whether a condition holds, choose from fixed options or place an input on an ordered rubric. Use this GPT-6 Luna mode when the task needs a defined decision result rather than a written response.

Shared input, named questions and answers

A producer request contains the model, shared input and a questions array. The input can be a text string or user messages with text and images. Each question has a unique name, a type, instructions and any required options or levels. The answer array echoes the question names, allowing the result to be associated with the condition, category or rating that was actually requested.

For a support message, supply the message and the context relevant to the next decision, then name the question around the field the workflow needs. For a document passage, include the text that establishes the criterion instead of an unsupported reference to a larger file. Define the question around observable facts and separate independent concerns. A useful result can combine an issue category with a severity rating while retaining the definition and meaning of each answer.

Predicate: estimating whether a condition holds

A predicate question returns a probability from 0 to 1 that the defined condition is true. It can check passage relevance or a visible condition in an image. For a product photograph, describe the visible defect to check, such as a crack or tear, and provide the image with that criterion. The value is the model’s estimate; the application chooses the threshold used for flagging or review.

OpenAI’s Decisions guide requires inline encoded image data for this endpoint. Hosted image addresses and file identifiers are not supported image inputs in that documented contract. Text and image content can be combined in a user message so that the visual evidence is evaluated with the relevant written context. Keep the image legible and the condition specific; a probability for one stated defect is not a general inspection report covering every possible issue.

Choice: selecting one supplied value

A choice question selects from options with distinct values and descriptions explaining when they apply. The returned choice is one of those values, with per-option probabilities and a separate confidence field. For routing, define categories around the actual teams or outcomes available, and include an other option when the set does not cover every input. Keep these categories unordered: their positions are not a severity scale.

A practical classification brief can distinguish a product defect, an access issue and a general request. Describe the boundary between those categories instead of relying only on short names. Use the returned choice for the proposed label and inspect the distribution when several meanings remain plausible. The result supplies a selection; the application determines whether that selection should trigger an automatic route or a review.

Score: a probability-weighted rubric position

A score question uses ordered levels arranged from lowest to highest. Level indices start at 0, and the score is the probability-weighted average of those indices. It can fall between levels. For issue severity, one level can mean cosmetic impact, another a usable workaround and another complete blockage; define the criteria that separate them. The answer includes the per-level probabilities and confidence along with the score.

The expected score summarizes the distribution rather than choosing a single category. Keep that distinction when using a result for prioritization or an aggregate trend. A fixed-category display can use a choice question instead. Confidence is a separate reported field for choice and score answers; the guide does not define it as identical to the selected option’s probability.

Several questions and dependent decisions

Independent questions can share input in one producer request, including different question types. A photo could support a defect predicate and a category choice, or a message a category and severity evaluation. Give each concern its own instructions and meaning. When a later decision depends on an earlier result, the producer recommends separate requests so the second question can use that result explicitly. Do not assume that ordering questions in one array makes them a sequential conversation.

Choosing the right output and interpreting limits

Decisions supplies the defined predicate, choice or score result. An answer can instead have the type refusal; check its type before reading a probability, choice or score. OpenAI distinguishes this from generating arbitrary schema-based objects or written explanations through Structured Outputs, and from requesting a tool action through function calling. A decision answer should not be presented as a generated justification or proof that an external action occurred. General chat-model output, reasoning and tool settings are not automatically part of this specialized mode.

Use representative labeled inputs to assess thresholds for routing, filtering and review, taking account of the consequences of false positives and false negatives. Keep the evidence, question definitions and raw numerical outputs associated so an ambiguous result can be inspected. Keep the shared evidence relevant to the named questions, and define each option or rubric level clearly before submitting the decision task. Prepare the evidence and answer types around the documented decision workflow, with the output interpreted according to the task it evaluates.