OpenAI Decisions API

Last updated:

WrapperProprietary

⚠️ Why not a model

Documented as gpt-6-luna behind a typed endpoint; no decision-trained weights, head or model card. GA since 22 Oct 2026, same model.

Technical specs

Base LLMs
GPT-6 Luna (gpt-6-luna)
Decision types
choice, score, boolean
Features
hosted-api, generally-available, predefined-answers, text-input, image-input, json-input, prompt-caching, public-docs, input-only-pricing
License
Proprietary

The OpenAI Decisions API was announced at OpenAI DevDay on 29 September 2026. OpenAI’s recap describes it this way: it “enables real-time decision-making by focusing Luna’s intelligence on a specific set of user-defined questions with finite pre-defined answers. Developers supply context using text or images, and get back answers they can use to classify content, route requests, or choose an agent’s next action.”

Update, 6 October 2026: public beta. OpenAI opened it to all developers (announcement, 17:53 BRT; changelog). The sections from “What OpenAI has said” down describe the 29 September preview and are kept as history. See news.

General availability (22 Oct 2026)

OpenAI made the Decisions API generally available on 22 October 2026 (blog; changelog; forum, 17:51 BRT). What changed since the public beta, per OpenAI’s guide:

GA (22 Oct 2026)
Model still gpt-6-luna only, on POST /v1/decisions
Questions per request up to 400 (was 200)
Input text, images (inline or hosted URLs), native JSON (input_json), images inside individual question instructions, conversation history with tool calls
Price $0.10 per 1M input tokens; cached input $0.01 per 1M; no output charges
Other prompt caching, native GPT-Live integration, US/EU inference residency, Private Link, usage reporting

OpenAI says it is “up to 10x faster than GPT‑6 Luna through the Responses API” and quotes customer numbers (Sierra, Speak); all UnverifiedClaim. Independent numbers so far: JevBench API board Intelligence 56.9 (Jev 1.13: 63.6); Artificial Analysis Decision Index 48.5 for “GPT-6 Luna (Decisions)” (Jev 1.14: 57). See news.

Still a runtime. GA documents the same model behind a typed endpoint, with no decision-trained weights or decision head, so the page stays in (NOT)Models.

Public beta (6 Oct 2026)

Published by OpenAI
Endpoint POST /v1/decisions; gpt-6-luna is “the only model currently available”
Question types predicate (probability that a condition is true), choice (one of your options, with probabilities over the options), score (probability-weighted average over ordered level indices)
Input Text, images or both; several named questions per request
SDKs Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0, Java 4.78.0 or later
Price $0.10 per 1M input tokens; no output, cache-read or cache-write charges; long-context and regional multipliers apply
Status Public beta; GA expected “in the coming weeks” (guide)

Independent result. JevBench API board (v1.6.1 scale, revision v1.7.10, 6 Oct): Composite 62.5 (#5), 95% interval 51.6–64.6; Capability 73.5 (#4 among Jev-class APIs); Intelligence 56.9, Calibration 90.1; $0.052 per 1,000 decisions (1.6× Jev); raw p50 0.30 s. Measured on a fresh sealed set (A5) plus the public items and equated to the full-set scale.

Still a runtime. The docs describe a typed endpoint over gpt-6-luna, not decision-trained weights or a decision head, so it stays here and off the quadrant. “10x faster than the Responses API” is OpenAI’s claim (UnverifiedClaim).

Until 6 October it was in limited preview: access was “limited to selected API customers for testing”.

Name clash. OpenRouter already runs a different product called the Decisions API (POST /api/alpha/decisions), which serves Jev and other decision models. On this site the OpenAI product is always “OpenAI Decisions API”.

What OpenAI has said

Published by OpenAI (29 Sep 2026)
Announced DevDay 2026 recap; @thsottiaux at 14:27 BRT; @OpenAIDevs at 15:34 BRT
Engine “Powered by GPT-6 Luna” (OpenAIDevs); “focusing Luna’s intelligence” (recap). The Decoder reports “a version of GPT-6 Luna”
Input Text or images as context
Output “A selection your app can use” (OpenAIDevs follow-up) from answers you define
Uses named Classify content, route requests, choose an agent’s next action
Access Limited preview for selected API customers; broad release “in the coming days”
Docs / endpoint / schema None public on 29 Sep 2026. No changelog entry, no guide, no API reference, and no decisions endpoint on the GPT-6 Luna model page
Pricing Not published. GPT-6 Luna’s token prices ($0.10 / $0.50 per 1M) belong to Luna on Responses and Chat Completions; nothing says the Decisions API uses them

An OpenAI spokesperson told The New Stack that the company plans to share more “at broad rollout”.

What we don’t know yet

  • Endpoint path, request and response schema, and SDK support
  • Whether a response carries probabilities or a confidence value, whether either is calibrated, and whether there is an abstain option
  • Whether one call can answer several questions, and whether there are yes/no or ordinal (score) types beyond picking one answer
  • Maximum number of answers, context limit, regions and data residency
  • Latency distribution (p50 / p99), rate limits and pricing
  • Whether OpenAI trained or tuned decision-specific weights, or serves stock Luna with constrained output

Why it is listed here (not in Models)

The catalog admits models with trained decision weights and at least a minimal public API contract (criteria). As of 29 September 2026 the OpenAI Decisions API has neither on record. OpenAI describes it as a way of using GPT-6 Luna, a general text model, and has published no model card, no decision-specific weights and no API documentation. That makes it a hosted decision interface over an existing LLM, which is what this section covers. We list it as wrapper because that is the closest kind the schema offers.

Re-evaluation trigger: at broad release. If OpenAI documents tuned decision weights or a decision head, publishes a typed schema with probabilities, and puts out enough docs to score it, it moves to the catalog and the quadrant. Until then it stays off the quadrant.

status: active is the closest value the runtimes schema allows. The product is a live preview, not generally available.

Vendor numbers (UnverifiedClaim)

  • 150 ms vs 1.6 s per request, “~10x faster decisions”. From the 29-second launch video: a demo routes “10,000 customer requests” from customer_requests.csv to Billing / Technical / Sales, side by side with the Responses API. The video is marked “Shown 15x realtime” and shows no code, endpoint, model name or probabilities. The Decoder and The New Stack repeat the same figures.
  • “Less than a few hundreds of milliseconds end to end.” Thibault Sottiaux (OpenAI), who also says it “supports visual inputs”.
  • “Confidence scores.” Press wording (The New Stack), not OpenAI’s. OpenAI’s own posts only say the API “returns a selection”.

None of these come with a latency distribution, input size, region or SLA. No benchmark has been published on typed-decisions, JevBench, S1Bench or anything else.

Relation to System One and Jev

OpenAI does not use the “System One” name, does not expose /v1/systemone, and has not published Choice / Noul / Score types. Nothing public makes it compatible with TypeSafe’s API. The link to Jev comes from the timing (two weeks after Jev’s launch) and from press framing: The New Stack headlined it “OpenAI answers TypeSafe’s Jev”, and The Decoder says OpenAI “is moving into the field of so-called System One models”.

TypeSafe founder Diogo Almeida (@CompleteSkeptic, 15:24 BRT) posted a “clone war” meme, then: “jk, I love openai and think more competition and validation is great for developers! (assuming the model is good - plz make it good!) hopefully this is a sign for the future that building in a system one compatible way is the future.” The same day he linked TypeSafe’s system-one-adapter-python (MIT), a drop-in replacement for the TypeSafe client backed by LLM APIs, OpenAI included, for comparing TypeSafe against an LLM. It uses regular LLM endpoints, not the Decisions API.

Artificial Analysis Decision Index v1.0 (13 Oct 2026)

“GPT-6 Luna (Decisions)” appears among the decision models in the AA launch article charts; the text gives no figures for it, so none are quoted here. (GPT-6 Luna (Low) as an LLM scores 55 on the same index.) AA is a separate, independent board (four evaluations: AA-Omniscience-Decision, HLE-MultiChoice, When2Call, WildGuardTest). Its scores and ranks are not comparable with Decision Index 0.3 or JevBench. AA finds every decision model overconfident overall, mainly on the knowledge and reasoning tasks; most are fairly well calibrated on When2Call and WildGuardTest.