OpenAI’s Decisions API Reaches General Availability with GPT-6 Luna
OpenAI has made the Decisions API generally available. It uses GPT-6 Luna to return typed probabilities, category choices, and rubric scores for text, images, JSON, and conversation history.
#openai-api#llm#ai-development#classification#model-evaluation
OpenAI has made its Decisions API generally available, moving the endpoint beyond public beta. Requests go to POST /v1/decisions and use gpt-6-luna, the only model currently supported for the API.
The endpoint is designed for evaluations with a defined answer shape: whether a condition is true, which option from a supplied set fits, or how an input rates against ordered levels. OpenAI says Decisions returns these typed answers about 10 times faster than the Responses API. That is OpenAI’s performance claim, not a guarantee for every workload.
A focused endpoint for evaluation
A Decisions request has three main parts: a model, shared input, and questions. The input supplies evidence for the questions, while each question specifies its type, instructions, and, where relevant, choices or score levels. Each question needs a unique name, which the API echoes in its answers array.
There are three question types:
- Predicate estimates the probability, from 0 to 1, that a condition is true. For example, an application could check whether a product photo shows visible damage.
- Choice selects among supplied values, such as content categories or departments. The options have no required ordering.
- Score evaluates an input against ordered levels, such as issue severity. It returns probabilities across the levels and a probability-weighted average of their numeric indices, so the result can fall between levels.
That fixed structure is useful when an application needs an evaluation result to drive its own logic, such as routing a request or prioritizing work. It does not replace every structured model response. OpenAI recommends Structured Outputs with the Responses API when the task is to produce an object matching a custom JSON schema, such as extracted fields or an explanation. Function calling is the option when a model needs to request a tool call with arguments.
What changed since beta
The generally available version adds several capabilities. Developers can reuse shared input with prompt caching, attach images to individual question instructions for comparisons against shared evidence, and pass structured data directly using an input_json content part. A request can now contain up to 400 questions, up from 200 in the public beta.
Image inputs can be inline base64 data URLs or publicly accessible HTTP and HTTPS URLs. Decisions can also evaluate conversation history, including messages, function or custom tool calls, and their outputs. Tool outputs must be associated with their calls using call_id.
The input format has boundaries. User messages support text, images, and JSON; developer and system messages support text, with system messages treated as developer messages. Assistant messages can contain output text or refusals, and tool outputs can contain text and images. Files, audio, and item references are not supported. JSON content parts are supported only in user messages; keys inside the JSON are treated as data, not as instructions to create messages or fetch images.
Other additions include organization usage reporting for token use and request counts, native GPT-Live integration, and Private Link for sending requests over a private connection. With GPT-Live, Live can call Decisions directly or trigger decisions as input or output transcripts arrive.
Choosing where it fits
The key design question is whether a task has a clear set of answer types. A predicate fits a yes-or-no condition where a probability is useful. Choice fits selection from a defined list. Score fits ordered levels where the relative position matters. Teams should define instructions and allowed options carefully, then decide how their application will handle low-confidence or refused answers; the API returns typed results, but the application runs any actions those results select.
The Playground offers a way to experiment with questions and inputs before writing code. For production integrations, the documented request shape also supports combined evidence, such as text alongside an image or JSON object, without requiring every input to be flattened into a text string.
Bottom line
General availability makes Decisions a defined API option for applications that need model-based classification, routing, or scoring in a typed format. GPT-6 Luna is currently the sole supported model. Developers should check that their evidence fits the supported input types and choose Decisions for evaluations, rather than using it as a substitute for schema-based generation or tool calling.
Sources
Found this useful?
Share it with someone who follows open source.
Written with the help of AI from the sources listed above and OSRepos data, and reviewed before publishing. How we measure projects.