Use the OpenAI Decisions API for fast typed routing and scoring

The OpenAI Decisions API is in public beta, released on October 6, 2026. The current model is gpt-6-luna, through POST /v1/decisions. OpenAI's announcement says it is available to all developers; that is not a promise of free credits, every geography, unrestricted accounts or unlimited quota. Check the access and limits that apply to your API organization. Release changelog; official announcement.

For a fast classification API, first define the answer your code consumes. A known department, condition estimate or ordered severity fits Decisions. A narrative answer, extracted object with your own fields or model-requested tool call points toward Responses. The Decisions guide makes this distinction; the Luna model page documents Responses tool support.

Speed claim: OpenAI says Decisions using GPT-6 Luna can be up to 10x faster than GPT-6 Luna through the Responses API. This compares API paths; it does not establish that every request is exactly ten times faster or that Luna universally outperforms another model. We have not benchmarked either path. The table below is an editorial selection aid, not a measured comparison.

A request contains model, input and questions. Give questions distinct name values so the corresponding answers are easy to identify. Input can be a text string or user messages containing input_text and input_image parts. Images may use Base64 Data URLs or publicly accessible HTTP(S) image URLs. Private image URLs or URLs requiring authentication are not confirmed as supported. The reference allows up to 128 images across one request.

File IDs (file_id) are still not supported. Other message roles, function calls, files, audio, tools and item references are outside the documented Decisions input contract. A separate voice integration does not turn this endpoint into an audio-upload API. Request reference.

One shared input may have multiple independent questions in its questions array, including different question types. If a second decision depends on the first answer, separate the requests as the guide recommends. Combining related decisions is not a demonstrated accuracy improvement or blanket cost saving. Multiple-question guidance.

A predicate answer contains type, name and probability, an estimate from 0 to 1. For example: Does this support request mention a billing problem? The probability describes the model's estimate of that condition. It is not a certified probability or guaranteed truth. Question and answer types.

Original risk-triage example — not executed: a request says, “Please change the destination on the order; I cannot confirm which address is on file.” Ask: Does this request need human review because the requested change is ambiguous or requires identity verification? Use the estimate to prioritize a review queue. It must not authorize an address change, account action or payment.

Make the condition observable and narrow. “Is this a good customer?” mixes undefined judgments; a specific condition about missing information is easier to label and evaluate. A refusal must enter a separate handling path rather than becoming probability zero.

A choice question takes developer-supplied options. Supply 2–255 unique choices, as required by the current reference. The answer returns choice, probabilities and confidence, alongside type and name. Each probability is associated with an option value. Define values and descriptions that are understandable, distinct enough to route reliably and broad enough to cover the intended requests. Choice reference.

Original support-routing example — not executed: “My receipt lists the wrong billing address, but the parcel tracking looks correct.” Use billing for invoice/accounting questions, technical for product-function problems, shipping for parcel movement or delivery, and other for requests outside those meanings. In a manually labeled example, billing would be the intended queue; that is an editorial label, not an API prediction we obtained.

Document how to handle messages spanning two queues. The selected option does not remove that ambiguity. Inspect the distribution and validate your review rules on labeled requests. Confidence is a decision signal, not a probability of factual correctness or permission to act.

A score question defines ordered levels, each with a label and an optional description. Its answer contains score, per-level probabilities and confidence, plus type and name. Level indices begin at 0. The score is the probability-weighted average of those indices, so it can fall between integer levels. It is not a selected single label. Scoring guide.

Original severity example — not executed: a support dashboard has a broken saved-view filter, but users can still search manually. Define low (0) as a display inconvenience with the task intact; medium (1) as an impaired task with a workable alternative; high (2) as a task blocked without an available alternative. This rubric addresses queue priority, not a medical, legal or security risk assessment.

Illustrative arithmetic, not a model result: probabilities of 0.20, 0.50 and 0.30 yield 0 × 0.20 + 1 × 0.50 + 2 × 0.30 = 1.10. The same average can hide different distributions. Review the probabilities as well as the score, and use choice if your application needs one category rather than an average.

This single illustrative curl example has not been run. $OPENAI_API_KEY is a placeholder; no key is included or needed to read this guide. Running it yourself would make a billable API request under your account's applicable terms. The synthetic text avoids customer records.

curl https://api.openai.com/v1/decisions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "input": "The receipt shows an old billing address; parcel tracking is correct.",
    "questions": [{
      "type": "predicate",
      "name": "billing_mentioned",
      "instructions": "Does this support request mention a billing problem?"
    }]
  }'

After a successful HTTP response, inspect the answers array and branch on each answer's type: predicate, choice, score or refusal. A refusal carries the question name (or null for an unnamed question), without a usable scored answer; other questions may still receive answers. Route a refused or unexpected answer to a safe fallback or review. Do not interpret HTTP acceptance as a usable business decision. Also handle request errors separately. Answer reference.

AgentSkillsHub editorial operational guidance: decision → threshold → ambiguous band → human review → action. Define which outcomes may proceed, which stay unresolved and who approves the final action. This is a design pattern, not an OpenAI-mandated numeric threshold.

Before relying on the decision, use labeled representative examples from the intended task. Include ambiguous, multi-topic and out-of-scope inputs. Check false positives, false negatives, the confidence distribution and behavior near each threshold. Choose review boundaries according to the consequences of mistakes, then inspect cases crossing those boundaries. The official guide recommends task-specific labeled examples; this page reports no benchmark or measured accuracy.

For payments, account suspension, employment, medical or legal decisions, and security-sensitive irreversible actions, do not use a single model score as the sole basis for automatic execution. A high confidence value cannot supply identity verification, organizational permission, professional judgment or an audit of the underlying evidence. Keep the action layer responsible for those checks.

As checked on 2026-10-07, the GPT-6 Luna Decisions endpoint costs $0.10 per 1M input tokens. Only input tokens are charged: no cache-read, cache-write or output-token charges. Regional processing premiums and long-context input pricing multipliers apply. These terms are for /v1/decisions; other Luna requests follow their own model and processing-tier pricing. Endpoint pricing; model pricing.

The current account tiers are Free, Build, Launch and Grow. Free is subject to allowed geography and lists a $100 monthly usage limit. The three paid tiers qualify on cumulative credit purchases: Build requires $5 and lists a $500 monthly usage limit; Launch requires $100 and lists $5,000; Grow requires $500 and lists $200,000. Organizations upgrade as purchase thresholds are met. Current tier table.

A monthly usage limit is not a monthly bill, included credits or a per-model rate limit. “Free” is not a promise of free Decisions credits. Check Settings → Organization → Limits for your effective RPM/TPM and model access. We did not inspect or change an organization's limits, tier or billing configuration.

Remove secrets, raw credentials and unnecessary sensitive data from the input. Images may contain private names, account details or other identifying information even when the visible task concerns an object. Send only the evidence needed for the question and keep operational logs appropriately limited.

Do not assume Decisions data is never retained. The current API data controls documentation lists abuse-monitoring retention for /v1/decisions and describes exceptions and eligibility for Zero Data Retention. Availability of a control does not mean it is active for your organization; confirm the current requirements, configuration and limitations before using sensitive inputs.

Next step: write one bounded routing question and its intended human-review rule, label a representative set of examples, and inspect the official Decisions guide before authorizing an experiment. Browse the workflow guides if your task instead needs an agent or MCP integration.