Limited preview · announced 29 Sep 2026 /v1/decisions live on Jev 1.13

OpenAI Decisions API, explained — with a decisions endpoint you can call today

OpenAI's Decisions API returns one answer from a finite list you define — built to classify, route, and pick an agent's next action. It's in limited preview with no public endpoint yet. This site tracks what OpenAI announced, and the playground below runs the same pattern today on TypeSafe's Jev System One model.

Independent resource — not OpenAI. The OpenAI model option is shown disabled in the model picker until OpenAI opens its API.

Live playground callable today
context · state0/8000
questions · finite answers
Result

Run a decision. Every allowed answer comes back with its probability — one is picked.

⌘/Ctrl + ↵10,000 free trial tokens · no signup
What OpenAI announced

Finite answers from a dedicated model — not a chat completion

Announced at OpenAI DevDay on 29 Sep 2026. Quotes below are from OpenAI's recap; speed numbers are third-party reported claims, not benchmarks we ran.

What it is

“Decisions API 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 as text or images and “get back answers they can use to classify content, route requests, or choose an agent's next action.” OpenAI DevDay recap

Model

Powered by GPT-6 Luna, per OpenAI's developer account. The Decoder reports OpenAI claiming ~10× faster than asking Luna via chat, and 76/78 correct steps in a computer-task simulation at ~230 ms per call. The Decoder

Latency & output

byteiota reports ~150 ms typical latency and a response that is one answer from your list plus a confidence score. Treat as reported claims until OpenAI publishes numbers. byteiota

Status & pricing

“Available in limited preview today with a broad release planned in the coming days.” No public endpoint, request schema, rate limits, or pricing are published yet. Full guide →

How a decisions call works

Context in, one answer out

Same shape whether it ends up calling OpenAI's endpoint or ours: you bound the answer space, the model scores it.

What it's for

Decisions your code can act on directly

CLASSIFY

Content classification

Moderation labels, intent tags, sentiment buckets — a fixed taxonomy in, one label per item out, with probabilities for the rest.

ROUTE

Request routing

Send each inbound request to the right queue, handler, or model tier. Replace a tree of regexes and prompt-parsing with one call.

AGENTS

Agent next action

The use case OpenAI demoed: give the agent a finite action space — answer, ask, call a tool, escalate — and get the next step.

CONTROL

Real-time control loops

Reported ~150–230 ms calls make per-request decisions plausible: feature flags, throttles, UI branching, and game or robot state.

Landscape

OpenAI Decisions vs what you can run today

“Not published” means the vendor has not published it — we don't fill gaps with guesses.

OptionStatusOutputReported speedPublished pricingCall it now?
OpenAI Decisions APIGPT-6 Luna · announced 29 Sep 2026Limited previewOne of your answers + confidenceschema not publishedNot published~150–230 ms reported by third partiesNot publishedWaitlist/preview only
/v1/decisions hereTypeSafe Jev 1.13 · System OneLiveAnswer + per-option probabilitieschoice · noul · scoreNot publishedplayground shows per-call latency$0.042 / 1M input tokens upstream; packs here from $9.90Yes — playground above
Structured outputs / JSON modeany chat model + schemaThe old patternJSON you must parse & constrainprobabilities not nativeVaries by modelVaries by modelYes, with extra code
Integrate

One POST, a bounded answer back

Hits this site's /v1/decisions (Jev 1.13 today). Keep your request shape — state + questions with finite answers — and swapping to OpenAI's endpoint later should be a base-URL change, not a rewrite.

curl https://opendecisionsapi.com/v1/decisions \
  -H "Authorization: Bearer $OPENDECISIONS_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ticket-4821-route" \
  -d '{
  "model": "jev-latest",
  "state": "Since yesterday, every CSV export from our dashboard stops halfway through...",
  "questions": {
    "route": {
      "type": "choice",
      "instructions": "Which team should handle this support ticket?",
      "criteria": {
        "Billing": "Invoices, payments, subscriptions, or refunds.",
        "Technical": "Errors or product features that are not working.",
        "Account": "Login, access, or account settings."
      }
    },
    "urgent_today": {
      "type": "noul",
      "instructions": "Does this need attention today?"
    }
  }
}'
Pricing

OpenAI hasn't published Decisions pricing. Ours is boring and public.

Input tokens only — output is free. Guests get 10,000 trial tokens; a free account gets 100,000. Packs from $9.90 / 15M tokens, subscriptions from $29/mo. What we know about OpenAI's pricing →

FAQ

OpenAI Decisions API questions, answered honestly

Is this site run by OpenAI?

No. OpenDecisionsAPI is an independent developer resource. We cover the OpenAI Decisions API because it is new and sparsely documented, and we run a working decisions endpoint on TypeSafe's Jev model so you can build against the pattern today. We are not affiliated with or endorsed by OpenAI or TypeSafe.

Can I call the real OpenAI Decisions API here?

No — OpenAI has released it only as a limited preview with no public endpoint. The OpenAI option in our model picker is disabled on purpose, and any request naming it returns a 422. When OpenAI publishes an endpoint and schema we will link it and, if practical, support it.

What does the Decisions API actually return?

Per OpenAI's DevDay recap: one answer from the finite set you define, which you can use to classify content, route requests, or choose an agent's next action. byteiota reports the response includes a confidence score. OpenAI has not published the response schema yet.

How fast is it?

OpenAI has not published latency numbers. Third parties report ~150 ms typical (byteiota) and ~230 ms per call in a computer-task simulation where it completed 76/78 steps correctly (The Decoder). Treat these as reported claims, not benchmarks.

What does the live model on this site do differently?

Jev 1.13 (TypeSafe System One) answers the same kind of bounded questions — choice, yes/no (noul), and score — and returns per-option probabilities plus a confidence for choice questions. It charges input tokens only.

What do I get for free?

Guests can run the playground with 10,000 trial input tokens. A free account gets 100,000 input tokens and API keys for /v1/decisions. Paid packs start at $9.90 for 15,000,000 input tokens and never expire.