# OpenAI Decisions API: Typed Answers, 10x Faster > Source: > Published: 2026-10-18 11:08:12+00:00 OpenAI shipped an endpoint that doesn’t generate text. The [Decisions API](https://developers.openai.com/api/docs/guides/decisions) — now in public beta — takes a question with a fixed answer set and returns a typed choice, probability, or score in roughly 150 milliseconds. That is its entire job. No prose, no chain-of-thought, no parsing required. The pitch is compact: stop paying for words your code immediately throws away. ## How It Works The endpoint is `POST /v1/decisions`, powered by `gpt-6-luna`. You send context (text or images) alongside a `questions` array. Each question has a type: - **Predicate** — returns a probability (0–1) that a condition is true - **Choice** — picks from a fixed list of options, returns the selection and a confidence score - **Score** — rates the input against ordered levels (Low / Medium / High / Critical) One request can bundle multiple questions and return all answers in a single call. Here is a predicate example: ``` decision = client.decisions.create( model="gpt-6-luna", input="Customer cancelled Sept 28 but was charged $120 on Oct 1.", questions=[{ "type": "predicate", "name": "refund_needed", "instructions": "Does this customer need a refund?" }] ) # answer.probability = 0.98 ``` The same call structure handles support ticket routing, agent action selection, incident severity scoring, and image-based return inspections. The output is typed, so your orchestrator consumes it directly without a parsing step. ## The 10x Speed Claim — What It Actually Means OpenAI says the Decisions API is roughly 10x faster than the Responses API. That comparison is real: around 150ms versus 1.6 seconds when both use Luna on a comparable classification task. The number in the launch announcement is accurate. It is also somewhat selective. TypeSafe Jev — the category pioneer — runs at 70–500ms end-to-end. [AWS Strands Decider 2B](https://superpowerdaily.com/posts/aws-releases-an-open-ai-model-that-chooses-an-agent-s-next-step-instead-of-generating-any-text), released October 1, returns a decision in a median 115ms on a consumer RTX 3090 and is free to self-host under Apache 2.0. Cloudflare shipped Clef this month. The Decisions API is 10x faster than OpenAI’s own Responses API — not 10x faster than its actual competition. That is a meaningful distinction if you are evaluating whether to adopt it. ## The Marble Jar Problem A developer ran a controlled test: a jar with 5 red marbles, 3 blue, 2 white. True probability of drawing red: 50 percent. With red listed first in the choices array, the API returned 86 percent red. With red listed last: 73 percent. That is a 13-point swing from answer order alone. A separate test with a coin biased to 70 percent heads saw the choice format return 98 percent heads; predicate format returned 70 percent correctly. The community’s read on this, documented in the [marble jar test](https://dev.to/dishant0406/the-openai-decisions-api-hype-met-a-jar-of-marbles-1jjp), is worth taking seriously: a model whose confidence shifts 13 points because you reordered a list is giving you a vibe, not a measurement. The API works well for routing. It should not be trusted as a calibrated probability instrument. Always include a fallback option — an “other” or “escalate” choice — and route low-confidence answers to human review. ## How It Stacks Up Against the Alternatives The decision model space filled up fast in October 2026. The three most relevant options for developers: - **OpenAI Decisions API** — Managed hosting, image support, tight integration with the OpenAI platform. Input costs $0.10 per million tokens. Still in limited preview; standard keys return 403 errors. No cached input tokens, which limits cost efficiency on repetitive classification tasks. - **[TypeSafe Jev](https://www.valyu.ai/blogs/jev-vs-decisions-api)** — Text-only, $0.042 per million tokens, complete documentation including published failure modes. Available now. The category original. - **AWS Strands Decider 2B** — Self-hosted, free, Apache 2.0. Training data and scripts on GitHub and Hugging Face. Best choice if you want zero marginal cost at scale. The practical split: images point you to the Decisions API. Cheapest hosted text decisions point you to Jev. Full stack ownership points you to Strands Decider 2B. ## If You Don’t Have Access Yet Standard API keys were returning 403 errors for two weeks post-launch. If you are on the waitlist, you can replicate most of the Decisions API behavior today using Luna with strict JSON schemas and `reasoning: none` through [Structured Outputs](https://platform.openai.com/docs/guides/structured-outputs). Migration to the Decisions endpoint later requires minimal code changes — the questions array maps cleanly to the JSON schema approach. There is no reason to wait on the waitlist before building. ## Bottom Line The Decisions API fills a real gap: typed, fast answers for agent control planes. The speed is genuine. The image support is a clear advantage over Jev. The launch execution — 403 errors at launch, no separately published pricing, speed claims benchmarked against OpenAI’s own slower endpoint — leaves something to be desired. More importantly, the calibration issues with choice ordering mean you should treat confidence scores as routing hints, not measurements. Use it for classification and action selection. Build in escalation paths. Test against your own data before trusting the numbers. The [OpenAI DevDay 2026](https://www.infoq.com/news/2026/10/openai-devday-2026/) recap covers the full platform context if you want more.