OpenAI Decisions API: How to Use

OpenAI Decisions API is in public beta for typed yes/no, choice, and score answers via gpt-6-luna at $0.10 per 1M input, with no output token charges.

OpenAI’s Decisions API is now in public beta. It is built for one job: turn text and images into typed answers your app can route on — not for long free-form generation. OpenAI says answers come back about 10x faster than the same model through the Responses API. This guide covers what it is, how to call it, verified pricing and when to use it instead of Structured Outputs.

What the Decisions API does

A request has three parts: model (currently only gpt-6-luna), input (shared evidence: text and/or images) and questions (what to evaluate). The response returns an answers array. Give each question a unique name so you can match answers in code.

Question type	Use it to	Main result
predicate	Check a condition (damage, relevance, policy match)	probability from 0 to 1
choice	Pick one option from a fixed set (department, category)	choice plus probabilities and confidence
score	Rate against ordered levels (severity, quality)	probability-weighted score that can fall between levels

Images on this endpoint must be inline base64 data URLs. Hosted HTTP image URLs and file_id inputs are not supported. You can ask several independent questions on the same input in one request; for dependent decisions, send separate calls.

OpenAI Decisions API pricing (2026)

OpenAI’s Decisions guide states that with gpt-6-luna, input costs $0.10 per 1M tokens. You pay only for input tokens: no cache-read, cache-write or output-token charges on /v1/decisions. Regional processing premiums and long-context input multipliers can still apply. We checked the guide on October 15, 2026.

Important: other requests that use gpt-6-luna outside Decisions follow normal model pricing on the OpenAI pricing page (for example $0.10 input / $0.50 output on standard short-context Responses). Do not mix those rates up with the Decisions-only “input only” rule.

How to use it

1. Endpoint: POST https://api.openai.com/v1/decisions
2. Model: gpt-6-luna
3. SDKs: OpenAI documents Decisions helpers in recent SDK versions (Python 3.26.0+, JS 7.30.0+ and peers).
4. Playground: try questions in the OpenAI Playground before wiring production thresholds.

Example uses OpenAI highlights: flag visible product damage from a photo (predicate), route a support ticket to billing vs technical (choice), or score issue severity on a rubric (score). Set thresholds from labeled examples in your own data — OpenAI does not publish a universal cutoff.

Decisions vs Responses / Structured Outputs

• Use Decisions when you need a probability, a fixed choice or a rubric score to drive routing, filtering or prioritization.
• Use Structured Outputs on Responses when you need a custom JSON object, extracted fields or a written explanation.
• Use function calling when the model should request a tool call with arguments.

If you are wiring this into an agent stack, also see our OpenAI AgentKit page and the GPT-6 Astra vs Sol model comparison. For agent safety context, read Are AI agents safe?.

Note: this is OpenAI’s Decisions API — not Microsoft Decision-1.

Availability notes

OpenAI says Decisions supports Zero Data Retention and HIPAA for eligible customers, with data residency in the United States and Europe (EEA + Switzerland) under the usual data-control agreements. The product is still in public beta, with GA expected in the coming weeks per OpenAI’s docs.

Frequently asked questions