The OpenAI Decisions API is a fast interface for making bounded AI decisions. OpenAI made it generally available on 22 October 2026 for tasks such as classifying inputs, choosing an allowed next action and scoring records against a rubric. It is designed for structured judgments, not as a universal replacement for open-ended text generation.
News date: 22 October 2026. Research verified: 24 October 2026. The primary reference is OpenAI's general availability announcement. All speed, price and customer-performance figures are attributed to OpenAI or the customers quoted in its announcement, not to independent RKWS tests.
What is the OpenAI Decisions API?
The Decisions API accepts context from text, images or structured inputs and returns a specific answer type rather than a long conversational response. Developers define an answerable question and its constraints, such as whether a document is missing information or which department should receive a ticket.
The API uses GPT-6 Luna and targets applications that must make many relatively narrow decisions with low latency. OpenAI reports that it can be up to 10 times faster than asking GPT-6 Luna through the Responses API for equivalent bounded tasks. That is a vendor-reported capability, not an across-the-board benchmark guaranteed for every workload.
Which three answer types does the Decisions API support?
Each supported answer format addresses a different engineering need. Picking the right type is part of designing a reliable workflow.
| Answer type | What the API returns | Typical use case |
|---|---|---|
| Predicate | Estimated probability from 0 to 1 for a stated condition | Is a submitted image likely to show product damage? |
| Choice | A selected option from developer-supplied allowed choices, with associated probabilities | Route a ticket to sales, support, billing or human review |
| Score | A probability-weighted value across ordered rubric levels | Rate urgency, document completeness or lead quality |
A probability is an estimate, not proof. For high-impact workflows, calibrate thresholds against known examples and keep human review for uncertain or costly cases. Predicates and scores should not be treated as independent evidence that an input is true, lawful or safe.
What changed with the October 2026 general release?
OpenAI says the production release extends the earlier public beta with more input options and deployment features. Developers can supply native JSON, include images within instructions and submit up to 400 questions in one request, twice the prior limit. The release also adds prompt caching and native GPT-Live integration for voice applications.
- General availability: dedicated Decisions endpoint available for production integrations
- Structured data: native JSON as context alongside text and image inputs
- Batching: up to 400 questions per request, subject to relevant product limits
- Voice: GPT-Live integration for event-driven decisions during conversations
- Operations: prompt caching and regional processing options, including US and EU residency
OpenAI also states that Zero Data Retention is available to eligible customers. Availability can be account-specific, so developers should verify their organization's data settings rather than assuming residency or retention controls are automatically enabled.
How fast and affordable is the OpenAI Decisions API?
OpenAI's launch announcement quotes $0.10 per million uncached input tokens and $0.01 per million cached input tokens, with no output-token charge for Decisions. These are launch-date terms and may change. Actual costs depend on accepted inputs, cached context, request volume and applicable account conditions.
| Metric | Reported at launch | Interpretation |
|---|---|---|
| Model | GPT-6 Luna | Service uses a dedicated bounded-decision interface |
| Uncached input | $0.10 / 1 million tokens | Illustrative unit pricing, not a fixed monthly plan |
| Cached input | $0.01 / 1 million tokens | Applies when eligible cached context is reused |
| Output tokens | No output-token charges | Not equivalent to a promise that all infrastructure costs vanish |
| Speed | Up to 10× faster than same-model Responses requests | Test with your own data, prompts and latency targets |
The OpenAI developer announcement repeats the main capabilities and pricing. Any real cost estimate should use your expected monthly volume and the current official pricing page, not an unverified “average price per action”.
Decisions API vs Responses API: which should developers choose?
Use Decisions when the correct output fits a small set of predetermined outcomes. Use a general Responses API call when the application must create original explanations, write long content, reason through open-ended problems or generate complex structured artifacts.
| Application task | Preferred starting point | Reason |
|---|---|---|
| Ticket category or routing | Decisions | Known choices and measurable confusion matrix |
| Lead qualification score | Decisions plus review rules | Ordered rubric with auditable decisions |
| Draft a client proposal | Responses | Needs substantial generated writing |
| Generate a technical explanation | Responses | Open-ended content and context needed |
| Choose whether an agent should ask for clarification | Decisions | Bounded action options |
| Approve a payment or delete records | Human-controlled workflow | High-impact action should not rely on model judgment alone |
For broader foundations, see our MCP versus API guide and AI agent governance guide. A Decisions call is a specialized API building block, not an alternative to authentication, business rules or accountable approvals.
Where can an Indian software business use bounded decisions?
Developers building CRMs, websites and support platforms can start with high-volume decisions whose correctness can be tested objectively.
- Website enquiries: route contact forms to sales, customer service or spam review using controlled labels.
- Customer support: estimate whether a message is urgent and select the relevant workflow.
- Document handling: classify business forms or decide whether key fields appear to be missing.
- Product operations: score a product issue from photos or text for reviewer prioritisation.
- Voice assistants: choose whether to clarify, answer or transfer as a caller speaks.
- Agent orchestration: decide whether to continue, retry or hand off a step.
These are candidate applications, not verified case studies from RKWS customers. If your project handles personal data, retain appropriate consent, access rules and review paths; do not expose financial or medical records to an external provider without the necessary protections.
How should developers evaluate the OpenAI Decisions API?
A feature launch is not an evaluation. Build a small offline test before wiring decisions into production actions. The goal is to measure what happens when the service is wrong as well as when it is fast.
- Define one question with a bounded answer and clear examples of each category.
- Collect representative labeled examples across easy, ambiguous and rare edge cases.
- Record a baseline from deterministic rules or the existing workflow.
- Test predicate thresholds or permitted choices in OpenAI's Decisions Playground.
- Calculate accuracy, confusion patterns, latency at useful percentiles and total cost.
- Add human review for low-confidence or high-impact cases.
- Monitor drift and rerun evaluations after major data or prompt changes.
OpenAI quoted customer evaluations from Box, Datadog, Ramp, Sierra and Speak in its announcement. These are useful examples of measured use, but their improvements may not transfer to a different dataset or configuration. Our agent monitoring guide explains why production latency and error tracking remain necessary.
What are the important limitations and safety questions?
Bounded responses make integration easier, but do not guarantee perfect classification. Developers should plan for contradictory evidence, misleading documents, adversarial text, unexpected data formats and unavailable services.
- Keep hard security policies in deterministic controls rather than relying on an AI score alone
- Reject or quarantine malformed inputs, missing fields and out-of-domain data
- Use explicit fallback behaviour when confidence is low
- Validate outputs against the allowed schema before taking action
- Restrict tool permissions and protect sensitive information
- Log decision identifiers, accepted inputs and human overrides as permitted by policy
See our prompt-injection and permissions article for the risk of treating external content as trusted instructions.
Frequently asked questions about OpenAI Decisions API
Is the Decisions API generally available?
Yes. OpenAI announced general availability on 22 October 2026 following a public beta.
What does a predicate result mean?
It estimates the probability that a statement holds, between zero and one. It is not certainty or an automatic approval.
How much does the API cost?
At launch, OpenAI listed $0.10 per million input tokens, or $0.01 for eligible cached input, and no output-token charges. Check current rates before budgeting.
Can the Decisions API work with voice applications?
OpenAI announced native integration with GPT-Live for decisions during live voice interactions.
Should every application move to Decisions API?
No. It is best for bounded, measurable judgments. Long-form generation and complex explanations still require more general models and application logic.
Developer takeaway
Use Decisions API as a narrowly scoped decision layer, not as a replacement for system design. Start with a single measurable classification or routing task; evaluate accuracy, confidence and end-to-end latency before scaling. For help planning a web app or integration, see RKWS software development services.
Primary sources: OpenAI: Build more responsive apps with the Decisions API, 22 October 2026; OpenAI Developer Community general availability announcement.
Frequently asked questions
What is the OpenAI Decisions API?
It is a dedicated API for bounded, structured decisions: estimating a condition's probability, selecting among supplied options or scoring input against a rubric.
When did the Decisions API become generally available?
OpenAI announced general availability on 22 October 2026, after an earlier public beta.
What are the three Decisions API answer types?
The API supports predicate, choice and score outputs for classification, routing and rubric-based evaluations.
How much does the Decisions API cost?
At the 22 October 2026 launch, OpenAI listed GPT-6 Luna Decisions input at $0.10 per million tokens and cached input at $0.01 per million tokens, with no output-token charges. Verify current pricing before production use.
Can OpenAI Decisions API replace a full generative response?
Not for open-ended writing or complex explanations. The Decisions API is designed for fast bounded judgments, while a full model response remains appropriate when the application needs generated content.
