OpenAI Decisions API: 150ms picks on Luna, price not yet public

OpenAI's new Decisions API does one job: you give it a question and a closed list of answers, and it gives back one of them. No paragraph, no parsing. OpenAI announced it at DevDay on September 29 as a limited preview. It is the latest in a run of decision-model launches, and the first from a frontier lab you probably already pay.

What actually happened

Per OpenAI's DevDay recap and The Decoder's coverage:

• How it works: you define questions with a finite set of predefined answers, send context as text or images, and get an answer back.
• Model: a version of GPT-6 Luna, OpenAI's low-cost model. Luna lists at $0.10 input and $0.50 output per million tokens in the regular API.
• Speed: about 150 milliseconds, which The Decoder reports as roughly ten times faster than asking Luna the same thing through the normal API.
• Named use cases: classify content, route requests, choose an agent's next action.
• Availability: limited preview, with broad release "in the coming days."

What OpenAI has not said: the price. There is no public rate card yet, and no word on whether it bills per call, per token or per question. Do not assume it costs the same as a plain Luna call.

It lands in a crowded week. AWS open-sourced Strands Decider 2B, and TypeSafe's Jev started the category.

Why a decisions API matters for your business

Most of your agent's calls are decisions. "Is this a refund, a lead or spam?" "Does this order need a human?" "Which tool runs next?" Each one is a pick from a short list. Today most teams send those to a chat model and regex one word out of the reply. A decision endpoint removes the parsing step and the latency that goes with it.

150ms changes where you can put AI. At that speed a check fits inside a checkout flow or a live chat turn, not just a background job.

Pricing decides the winner, and it is blank. A hosted API means no GPU to run. A local model like Strands Decider means a fixed cost and no data leaving your box. Until OpenAI publishes a price, you cannot do that math. Build your routing behind one interface so you can swap the backend when the numbers land.

Measure accuracy on your own data. Pull 200 labeled tickets or orders, run them through the preview and your current setup, and compare. Speed is worth nothing if one pick in five is wrong.

Sources: OpenAI DevDay 2026 recap, The Decoder.