Evergreen Tools
|
← Back to AI Tools

OpenAI Decisions API

Turns "what happens next" into one low-latency, low-cost API call instead of a paragraph of generated text

Official Source

This page summarises an official announcement — check the vendor site for current availability and pricing.

Features

  • ✓ Three answer types: predicate returns a 0-1 probability, choice returns one of your options, score returns a rubric-weighted score
  • ✓ OpenAI says bounded decisions run up to 10x faster than GPT-6 Luna through the Responses API
  • ✓ Priced at $0.10 per million input tokens with no output-token charge, and $0.01 per million for cached reads
  • ✓ Up to 400 questions per request (200 during the beta), with image, native JSON and hosted image URL inputs
  • ✓ Native GPT-Live integration for decisions inside live voice sessions, plus Zero Data Retention and HIPAA for eligible customers

How to Use

  1. Get an API key on the OpenAI platform and confirm gpt-6-luna access
  2. POST to /v1/decisions with a question definition: type (predicate, choice or score), name and instructions
  3. Pass context as text, images or JSON, up to 400 questions per request
  4. Read the probability or choice back into your routing, classification or validation logic, then A/B it against real samples

FAQ

What is the Decisions API?

An OpenAI endpoint that moved from public beta to general availability on 22 October 2026. It runs on GPT-6 Luna, answers only bounded questions (yes/no, pick one of N, or score against a rubric) and returns structured answers instead of prose, aiming to cut latency from the 1-3 second range to roughly 100-200 milliseconds.

How is it priced?

OpenAI lists $0.10 per million input tokens, $0.01 per million cached input tokens, and no output-token charge at all. Regional processing premiums and long-context multipliers still apply, so check the pricing and caching guide for your case.

When should I use each answer type?

Use predicate for questions like "does this product show visible damage"; use choice when an agent must continue, retry or ask for clarification; use score for ordered rubrics such as damage severity, where you get the probability-weighted average of the level indices.

What real numbers has OpenAI published?

OpenAI's customer quotes say Datadog scores agent-generated shell commands with 40x faster p95 at comparable accuracy, Sierra runs it in live orchestration at under 280 ms p95 for about 3x speed and 90% lower cost per call, and Speak cut time-to-next-action from 1.4 seconds to 0.13 seconds. These are vendor-relayed customer claims, so re-test on your own samples.

Is it the same thing as TypeSafe's Jev?

Similar idea, different implementation. TechCrunch reported that OpenAI did not train a purpose-built decision model; it reconfigured GPT-6 Luna to speed up inference and cut cost, so you inherit a general model's capabilities rather than a from-scratch calibrated decision network.