OpenAI · Decision model · Last verified 2026-10-15

OpenAI Decisions API

BetaPartly verified

Dedicated POST /v1/decisions endpoint that returns typed answers (probability, choice, or score) about text and images instead of generated text. Model: gpt-6-luna (only model available).

What it does

  • Question types: predicate (probability 0–1 that a condition is true), choice (one supplied value + per-option probabilities + confidence), score (probability-weighted average over ordered levels, indices from 0).
  • Input: text string or user messages mixing input_text and input_image. Images must be inline base64 data URLs; hosted URLs and file_id are not supported.
  • Multiple independent questions can share one request; refusals come back as a typed "refusal" answer.
  • SDK minimums: Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0, Java 4.78.0.

Pricing

  • $0.10 per 1M input tokens; no output, cache-read, or cache-write charges. [Verified 2026-10-15 against official page]
  • Regional processing premiums and long-context input multipliers apply (amounts not stated on the guide). [Verified 2026-10-15 against official page]

Limits

  • Context window and rate limits are not stated on the Decisions guide. [Unverified — not published on the page checked]

Status

  • Public beta; OpenAI says it expects GA "in the coming weeks". [Verified 2026-10-15 against official page]
  • Data controls: ZDR and HIPAA for eligible customers; data residency/regional processing in the US and Europe (EEA + Switzerland). [Verified 2026-10-15 against official page]
  • No published SLA, dated model snapshot, or deprecation schedule found for /v1/decisions. [Absence noted 2026-10-14/15; not confirmed against enterprise terms]

Safety & fit notes

  • Good fit: classification, routing, prioritization, and review queues with app-owned thresholds set from your own labeled examples.
  • Only the alias gpt-6-luna is documented — behavior can change under you. Run a canary eval and log the model string on every call.
  • No rationale is returned; if users need an explanation, pair it with a generative model and human review.
  • OpenAI Usage Policies restrict automating high-stakes decisions in sensitive areas (e.g. critical infrastructure, product safety components, medical) without human review. Keep it advisory; never the sole authority.

Compare

Sources