Most AI APIs answer in words. You ask a question, the model writes a paragraph, and your code extracts the actual answer with a regex and a prayer.
OpenAI's Decisions API, in public beta since October 6, 2026, skips that step entirely. It generates no text. It returns a typed verdict: a probability, a choice from a list you define, or a score against a rubric you define.
Three question types
Predicate: one condition, one probability from 0 to 1. OpenAI's example inspects a product photo for damage and returns 0.92 — your code compares it against a threshold.
Choice: pick one from a fixed list. The complaint "I was charged twice" routes to billing with 0.95 probability and 0.93 confidence. Include an "other" fallback so uncovered cases land in a review queue.
Score: rate against ordered levels. Bug severity across Cosmetic (0), Workaround available (1), Fully blocked (2) with probabilities 0.1, 0.7, 0.2 returns a score of 1.1 — between levels.
Why it matters
A probability supports a policy: auto-approve above 0.9, escalate between 0.6 and 0.9, hand to a human below 0.6. A bare label does not. Structured Outputs guarantees the shape of an answer; Decisions tells you how sure the model is.
Pricing
$0.10 per 1 million input tokens with gpt-6-luna, and zero output charges — there are no output tokens. Classifying a 500-token support message costs about $0.00005.
Caveats
Beta-only, one model, no published benchmark for the claimed 10x speed, probabilities are estimates (calibrate thresholds on your own labeled data), images must be inline base64, and no chaining — dependent decisions need separate requests.
Verdict
If your app asks a language model yes-or-no questions today, prototype this week. If it needs explanations or custom JSON shapes, the Responses API stays home.
Read the full breakdown — ticket-routing example, Decisions vs Structured Outputs vs function calling, pros and cons — on NeurAITool: https://www.neuraitool.com/2026/10/openai-decisions-api-answers-in.html
Top comments (0)