OpenAI Decisions API: GPT-6 Luna Choices in Limited Preview

The OpenAI Decisions API picks from developer-defined answers and returns a confidence score for classification or routing. It was in limited preview, with no published price.

More from this month: September 2026 AI product and model launches

Questions with answers chosen in advance

The Decisions API is not a chat endpoint. A developer defines a question and a finite set of allowed answers, supplies context as text or images, and receives one of those answers back. OpenAI pitches it for classifying content, routing requests and choosing an agent’s next action, the kind of judgment that sits inside a larger program. [1][2]

OpenAI says the API focuses the intelligence of Luna, the smallest and cheapest tier of the GPT-6 family released a week earlier, on the developer’s question. The Decoder describes the underlying model as a version of GPT-6 Luna, and The New Stack reports that answers come back with confidence scores rather than prose. [1][3][2]

The speed claim

OpenAI’s headline figure is speed. According to The New Stack, OpenAI says a decision returns in 150 milliseconds, against 1.6 seconds for an ordinary GPT-6 Luna call; The Decoder reports OpenAI’s framing as ten times faster than Luna through the regular API. Both are OpenAI’s own numbers, published with no benchmark method, and no independent measurement existed on launch day. [2][3]

A limited preview with the details missing

On September 29 the API was in limited preview, with OpenAI saying a broad release was planned in the coming days. An OpenAI spokesperson told The New Stack the company would share more at broad rollout. Price per call, the number of candidate answers a request can carry, and whether developers can tune it on their own data were all unpublished. [1][2]

Two weeks after Jev

Atlas interpretation: The announcement came two weeks after TypeSafe AI opened early access to Jev, a model built to return typed decisions with probabilities. The New Stack and The Decoder both placed OpenAI’s API in that category, and The New Stack judged it likely a rushed response. The idea of scored answers is not new at OpenAI: its Moderation API has long returned per-category scores, but with categories OpenAI fixed rather than ones the developer writes. [2][3]

Sources

  1. DevDay 2026 Recap

    OpenAI · Sep 29, 2026

  2. OpenAI answers TypeSafe’s Jev with a Decision API built on Luna

    The New Stack · Sep 29, 2026

  3. OpenAI expands Codex and its API at DevDay with security scans, a Decisions API, and Ultrafast

    The Decoder · Sep 29, 2026