October 14, 2026 · Ahmed · More by Ahmed
OpenAI's Decisions API Skips Generation for Routing and Scoring
OpenAI's new Decisions API returns typed probability answers instead of generated text — built for classification, routing, and scoring, not chat.
What the Endpoint Actually Returns
OpenAI's Decisions API, now in public beta, is a dedicated POST /v1/decisions endpoint that sits apart from the Responses API. Instead of generating text, it evaluates an input — a text string, or a set of messages mixing text and images — against one or more named questions and returns typed answers. As of this writing, gpt-6-luna is the only model it runs on. OpenAI says the endpoint returns answers "about 10x faster" than getting the same answer out of gpt-6-luna through the Responses API.
A single request can bundle several questions against the same shared input. Each question needs a unique name, and the response's answers array echoes that name back, so a caller evaluating multiple conditions at once — say, visible damage and a routing category from the same product photo — gets every answer matched to the question that produced it, in one round trip.
Three Question Types, One Subtle Distinction
There are exactly three question types. A predicate question returns a probability from 0 to 1 that a stated condition is true — useful for a yes/no check like "does this photo show visible damage?" A choice question selects one value from a fixed set of options and also returns a probability distribution across them. A score question evaluates an input against ordered rubric levels and returns the probability-weighted average of those levels' indices, rather than a single discrete pick.
That last distinction matters for anyone building a severity or triage scorer: because score takes a probability-weighted average, the number it returns can land between two defined levels instead of landing exactly on one, reflecting genuine uncertainty between adjacent tiers rather than forcing a hard pick.
Where OpenAI Draws the Line
OpenAI is explicit about what Decisions is not for. Its own guidance: use Decisions when the answer you need is a predicate, choice, or score; use Structured Outputs with the Responses API when you need the model to generate an object that follows your own JSON schema, such as extracted fields or a written explanation; and use function calling when the model needs to request a tool call with arguments. Decisions isn't a generation endpoint, and it isn't meant to replace either pattern — it fills a narrower gap: fast, typed answers for classification, routing, and scoring tasks that currently need a full generation call just to get a label or a number back.
Price and the Beta's Limits
Per OpenAI's public-beta announcement, gpt-6-luna pricing through the Decisions API is $0.10 per 1M input tokens, with no cache-read, cache-write, or output-token charges — a different structure from the token pricing that applies to gpt-6-luna on other endpoints. Prompt caching also isn't available for Decisions specifically yet; OpenAI staff have confirmed that's a beta-stage limitation, not a statement about where it's headed.