Orbital / OpenAI Decisions API: what businesses should test

OpenAI Decisions API: what businesses should test.

OpenAI's Decisions API is generally available for bounded classification, routing and scoring. Use this business framework to test quality, thresholds, cost and review.

AI workflows · Business evaluation

By Orbital Content Team · Published October 24, 2026

AI-assisted research and drafting with editorial review. Product, availability, pricing and data-control statements were checked against OpenAI's October 22 announcement and current developer documentation on October 24, 2026. The evaluation framework and business implications are Orbital analysis. Orbital did not call the API, reproduce OpenAI's speed claim, validate customer-reported results, test model quality, inspect infrastructure, audit data controls or review customer-specific terms.

OpenAI made its Decisions API generally available on October 22. The API takes text, images or structured JSON and returns bounded answers for questions such as whether a condition is true, which option from a fixed list fits best, or where an item falls on a defined rubric. OpenAI positions it for repeated, narrow decisions inside applications: classification, routing and prioritization rather than an open-ended conversation.

That narrower interface can make a workflow easier to measure, but it does not make the answer automatically correct or the resulting action safe. A company still has to define observable questions, label representative examples, choose thresholds from the cost of errors, review sensitive cases and monitor what happens after deployment.

What OpenAI released on October 22

The Decisions API is now generally available through POST /v1/decisions. OpenAI's current guide identifies gpt-6-luna as the supported model. A request can include shared text, images or JSON and between one and 400 questions. The result is typed instead of being a free-form paragraph.

OpenAI documents three answer types:

  • Predicate: an estimated probability that a stated condition is true.
  • Choice: one selected option from a fixed list, with a probability distribution and a separate confidence field.
  • Score: a probability-weighted average across defined rubric levels, plus the level probabilities and confidence.

OpenAI also documents prompt caching, hosted or inline image inputs, conversation-history inputs, organization-level usage reporting and native GPT-Live integration. The company says the endpoint is about ten times faster than using GPT-6 Luna through the Responses API. That is a vendor claim, not a guarantee for every payload, region, queue or complete business workflow; a buyer should measure end-to-end latency on its own cases.

Current documentation lists uncached input at $0.10 per million tokens for GPT-6 Luna and says output tokens are not charged. Cache writes and reads follow GPT-6 Luna caching rates, while regional-processing premiums and long-context multipliers can apply. This is the model-endpoint price, not the total cost of a working automation.

Separate documented facts from business implications

What the release supports—and what it does not establish
AreaDocumented statementPractical implicationNot established
AvailabilityThe Decisions API is generally available.A business can evaluate a supported production endpoint rather than a preview-only feature.Fitness for a particular workload, account quota, regional capacity or a customer's required service level.
InputsThe API accepts text, images and JSON; a request can ask up to 400 questions.Related checks can share evidence and structured application data can be passed directly.That combining many questions improves accuracy, reduces total latency or is cheaper for every workload.
OutputsPredicate, choice and score answers return probabilities, and choice or score includes confidence.The application can route cases with explicit thresholds and review bands.That a displayed probability is calibrated to the buyer's population or that confidence equals business certainty.
SpeedOpenAI reports about ten-times-faster responses than GPT-6 Luna through the Responses API.The endpoint is designed for low-latency decision points.Independent reproduction or equivalent end-to-end performance after retrieval, validation, approvals and downstream tools.
PriceOpenAI lists $0.10 per million uncached input tokens and no output-token charge for the Decisions endpoint.Direct model cost can be estimated from measured input and cache usage.Total implementation, review, retry, integration, observability or exception-handling cost.
Data controlsOpenAI says API data is not used for training unless a customer explicitly opts in to share data, lists no application-state retention for /v1/decisions, and describes default abuse-monitoring retention of up to 30 days.A buyer has a documented starting point for a data-flow review.That every input is permitted, the buyer is approved for Zero Data Retention, or contract and regulatory requirements are satisfied.
Reported outcomesOpenAI's announcement includes company- and customer-reported speed, quality and cost examples.The examples identify use cases worth testing.Independent benchmarks or results transferable to another company's data and operating process.

Choose a decision that is actually bounded

A strong candidate is repeated, has a fixed set of allowed outcomes, can be scored against known examples and does not require the model to invent policy. The API can return a useful signal; the surrounding system decides whether to automate, ask for review or stop.

Workload-fit screen
QuestionBetter candidateWarning signal
Are the outcomes predefined?Route an inquiry to one of six service queues.Invent a strategy or make an exception to policy.
Can the evidence be observed?Classify fields and language present in a submitted form.Infer private motives, protected traits or facts not in the input.
Can quality be labeled?Past examples have agreed outcomes or a repeatable review rubric.Reviewers regularly disagree and no decision owner resolves the rule.
Can errors be contained?Low-confidence cases become drafts for a trained reviewer.A wrong result sends money, deletes data, denies a right or creates a binding commitment.
Is the process stable?Categories, policy and inputs change through a controlled release.Rules shift informally and labels are never updated.
Is a narrower rule enough?Meaning depends on language or an image and rules perform poorly.An exact database lookup, validation rule or deterministic calculation solves the task.

Potential first tests include triaging a contact form, flagging whether a document has a required visual element, assigning an internal urgency rubric, or routing a support summary to the correct queue. Do not start with credit, employment, housing, medical, legal or other high-impact decisions without qualified policy, legal, privacy, security and domain review. Customer messages, payments, publishing, account changes and deletion should remain behind explicit approval until the business has evidence for safe automation.

Define the process first with Orbital's AI automation readiness checklist. The trigger, decision owner, allowed data, review path and fallback matter more than the API call.

Design answer types and thresholds around the business consequence

Start with one sentence describing the decision and one owner accountable for its policy. Then choose the smallest answer type that supports the action.

  1. Use a predicate for a gate. Example: “Does this inquiry explicitly request emergency service?” The output is a probability, so the workflow still needs thresholds.
  2. Use a choice for exclusive routing. Example: “Which service category best matches this request?” Make options distinct and include a safe “unclear” path when appropriate.
  3. Use a score for an ordered rubric. Example: classify operational impact as cosmetic, workaround available or fully blocked. Define each adjacent level with observable criteria.

Do not treat the highest probability as permission to act. Set three bands from labeled examples:

  • Automatic band: the observed error rate and business consequence meet a pre-approved threshold.
  • Review band: a trained person sees the input, answer, evidence and proposed next action before anything consequential happens.
  • Stop band: missing data, refusal, prohibited subject matter, conflicting checks or a threshold outside the accepted range sends the case to the manual process.

Thresholds are not universal. A false negative that delays an ordinary sales inquiry is different from a false negative that misses an urgent safety issue. Measure false positives and false negatives separately, assign a real operational cost to each, and choose the threshold before viewing live results.

OpenAI's guide explicitly recommends labeled examples and thresholds based on the cost of false positives and false negatives. A returned probability is an estimate from a model; it is not proof and should not be described to customers as objective certainty.

Build an evaluation set before building the integration

Collect representative cases from the intended process, remove or replace data the pilot is not approved to use, and have qualified reviewers assign the expected outcome. Include ordinary cases, difficult edge cases and examples that should stop rather than produce an automated action.

Minimum scorecard for a bounded decision
MeasureWhat to record
Class qualityPrecision and recall per outcome, not only overall accuracy.
Critical errorsAny prohibited route, missed urgent case, unsafe recommendation or unsupported action.
CalibrationWhether cases predicted near a probability level are correct at a similar observed rate.
Review loadPercent sent to humans, review minutes and queue delay.
Refusals and stopsCount, cause and whether the fallback worked.
LatencyMedian and tail time for the complete workflow, not just the model request.
ReliabilityTimeouts, malformed inputs, duplicate events, retries and dependency failures.
CostEndpoint, integration, monitoring and labor cost per accepted decision.
StabilityVariation across repeated cases and regressions after model, prompt, schema or policy changes.

Split the examples so threshold selection and final evaluation do not use the same cases. Version the questions, option definitions, examples, model and policy together. When any one changes, rerun the fixed evaluation before releasing it.

Compare against the current process. The baseline may be a deterministic rule, manual queue or another model. A faster answer is only better if accepted outcomes, review burden and downstream errors also improve.

Review data flow, retention and eligibility

OpenAI's current data-control table says data sent to /v1/decisions is not used to train models unless a customer explicitly opts in to share data, has no application-state retention and is eligible for Zero Data Retention subject to limitations. By default, abuse-monitoring logs may retain customer content for up to 30 days. Zero Data Retention and Modified Abuse Monitoring require eligibility, prior approval and additional requirements.

Prompt caching adds another detail: OpenAI documents that encrypted key-value tensors may be held in GPU-local storage and are not retained after a 24-hour expiration. Image inputs are scanned for child sexual abuse material; the documentation describes a retention exception when the classifier detects potential CSAM, including for approved data-control configurations.

Do not translate “ZDR eligible” into “our account has ZDR.” Verify the organization's project setting, written eligibility, applicable agreements, endpoint, model, region and every connected service. Data sent to CRM, analytics, logging, storage, support, observability or another processor follows that system's controls too.

Create a data map that records:

  • the source, owner and permitted purpose of every input;
  • personal, confidential, regulated and prohibited fields;
  • masking or minimization before the API request;
  • the OpenAI organization, project, endpoint and processing region;
  • prompt-cache use and whether image inputs are allowed;
  • application logs, traces, screenshots, queues and backups;
  • who can inspect content and operational metadata;
  • retention and verified deletion for each copy; and
  • third-party destinations reached after the decision.

Use Orbital's tech stack audit checklist to identify the systems, owners and duplicate data paths before adding another decision service.

Measure complete operating cost

The documented endpoint rate is unusually low, but token price can become the least important line item. Calculate cost per accepted decision, including:

  • uncached, cache-read and cache-write input tokens;
  • regional processing or long-context pricing adjustments;
  • retrieval, OCR, image preparation and upstream data cleanup;
  • queues, serverless compute, databases and network transfer;
  • authentication, secrets, logging, alerts and evaluation storage;
  • timeouts, duplicate events, retries and failed downstream actions;
  • human labeling, review, correction and exception handling;
  • engineering, security, privacy, legal and procurement work;
  • monitoring after model or policy changes; and
  • the manual fallback retained for outages or rejected cases.

Use organization usage reporting to reconcile tokens and requests by project or API key, then join that record with accepted outcomes and reviewer time. One request can contain many questions, so request count alone does not describe the amount of decision work or its quality.

Prompt caching may reduce input price when a long shared prefix repeats, but do not redesign a simple workflow only to maximize caching. Measure whether the common input is large enough, stable enough and safe to reuse; OpenAI's guide says the reusable prefix must contain at least 1,024 tokens.

Operate the decision as a controlled system

A production decision needs more than a prompt. Maintain a small control register:

  • business purpose, owner and technical maintainer;
  • question definitions, answer type and policy version;
  • model, endpoint, region and project;
  • allowed data classes and input validation;
  • automatic, review and stop thresholds;
  • actions the system may take and those it may never take;
  • human reviewer role, evidence shown and response time;
  • rate, concurrency and spending limits;
  • quality, latency, drift and exception alerts;
  • incident path, kill switch and tested manual fallback; and
  • last evaluation date and triggers for re-evaluation.

Log enough to reconstruct why the application chose a route without retaining unnecessary customer content. Preserve the model, question version, probability distribution, threshold, resulting action, approval and eventual outcome. Restrict access to this evidence and apply its own retention policy.

Watch the population, not only the model. Marketing campaigns, seasons, new services, spam patterns, form changes and customer language can change the input distribution. Alert when category mix, confidence, review rate or error rate departs from the evaluated range.

Keep irreversible or customer-impacting actions behind approval until measured evidence justifies a narrower boundary. An AI signal may draft a route; the application policy and accountable owner decide whether that route can run automatically.

A four-week bounded pilot

  1. Week 1 — define the decision. Select one internal, reversible workflow. Name the owner, fixed outcomes, prohibited cases, data boundary, baseline and success thresholds. Review current terms and project data controls.
  2. Week 2 — evaluate offline. Build labeled development and holdout sets. Write observable questions, compare predicate, choice or score formats where relevant, and select thresholds using false-positive and false-negative costs.
  3. Week 3 — run in shadow mode. Send permitted live inputs to the candidate without changing the real workflow. A human completes the existing process. Measure agreement, critical errors, review bands, latency, cost and failure recovery.
  4. Week 4 — release one reversible route. Enable one low-consequence action for a limited group and volume. Keep human approval for exceptions and customer-impacting work. Reconcile usage, audit logs and outcomes daily during the limited release.

End with a recorded decision: stop, revise, retain as a recommendation, or automate one approved band. Expand only one dimension at a time—traffic, data class, category or action—and rerun the holdout evaluation first.

Orbital's business take

The most useful part of this release is not that another model can classify text. It is that the product interface encourages teams to express an operational judgment as a bounded, measurable question with typed output. That can reduce fragile parsing and make threshold, review and monitoring decisions more explicit.

The risk is false simplicity. A low token price, probability and fast response can make a decision look production-ready before the business has defined its policy, labeled representative cases or measured downstream errors. The endpoint supplies a signal; it does not own the consequences.

A New York service business should start with one internal routing or prioritization task whose result can be checked and reversed. If the evaluation shows stable quality, lower total handling time and an acceptable review burden, the business has evidence for a controlled expansion. If it does not, the result is still useful: the team has learned where policy, data or process needs work before automating it.

Explore Orbital's AI automation services, tech stack consulting, connected-business planning, business consulting and service-business planning. A 60-minute consultation can turn one candidate workflow into a scoped evaluation brief; implementation, API usage, software subscriptions, security, privacy, legal review and ongoing support are separately scoped.

Sources and editorial limits

Primary OpenAI sources were checked October 24, 2026. OpenAI published the product announcement on October 22 and its documentation may change after publication. General availability, supported model, inputs, question limits, output types, pricing, regional processing, data residency and data controls are reported from OpenAI. Speed, quality and customer-result statements in the launch announcement are company- or customer-reported and were not independently reproduced by Orbital. Orbital did not call the API, compare it with another vendor, test probability calibration, validate ZDR approval, inspect OpenAI infrastructure, audit a customer environment or review customer-specific contracts. This article is business analysis, not legal, privacy, security, employment, credit, medical, financial or other professional advice, and it does not guarantee accuracy, speed, savings, compliance or business results.

Primary source links

Your next chapter.
Let’s give it lift.

Start a conversation