Last updated: 11/8/2026Last verified: 2026-10-23
What is OpenAI Decisions API?
OpenAI Decisions API is described in the supplied information as a developer API for bounded decisions over text, images, and structured input. Developers define predicate, choice, or score questions and receive probabilities, selections, or scores rather than free-form responses. Its stated applications include classification, routing, ranking, verification, and agent control. Product availability and technical specifications have not been independently verified for this review.
Key Features
- Bounded question types: The supplied description lists predicate, choice, and score questions for narrowly defined decision tasks.
- Multiple input formats: Stated support covers text, images, and structured JSON input.
- Decision-focused outputs: Listed outputs include probabilities, confidence, selections, and weighted scores instead of free-form text.
- Multi-question requests: The supplied specifications state a limit of up to 400 questions per request.
- Dedicated endpoint: The provided information identifies /v1/decisions as the API endpoint.
- Prompt caching: Listed as a supported capability, although cache behavior and eligibility are not specified.
- GPT-Live integration: Listed in the supplied information, but implementation details are not provided.
- Regional processing: Described as available to eligible customers, with regions and eligibility conditions unspecified.
Best For
Pricing
The supplied listing identifies OpenAI Decisions API as a paid product. It does not provide rates, billing units, minimum commitments, or trial availability. Confirm current pricing and access conditions through official documentation before estimating operating costs.
Pros & Cons
Pros
- The described output format is well aligned with workflows that need selections or scores rather than generated prose.
- Predicate, choice, and score questions offer different ways to express bounded decision tasks.
- The stated combination of text, image, and JSON inputs could support several types of application workflows.
- The listed multi-question request capability could help group related evaluations against shared input.
Cons
- The supplied information does not include benchmarks for latency, decision accuracy, or throughput.
- Pricing details are insufficient to estimate per-workflow operating costs.
- Probability and confidence outputs are listed without calibration evidence or guidance on interpreting them.
- Integration examples, error handling, rate limits, and detailed response schemas are not provided.
- Regional processing eligibility and GPT-Live integration details require clarification.
Alternatives
An alternative for developers building multimodal AI workflows with structured outputs and tool use. It is worth evaluating when an application needs broader generation capabilities alongside classification or routing.
Supports AI application workflows involving text and image analysis, tool use, and classification through developer-defined instructions. Its response formats and task performance should be evaluated against the application's requirements.
Offers multimodal models and structured output capabilities for applications such as extraction, classification, and routing. It provides another option for developers processing text and images.
A task-specific alternative for ranking documents by relevance to a query. It is relevant to the ranking use case rather than a replacement for every predicate, choice, or score workflow.
FAQ
Details
Platform
Features
- Answers bounded predicate, choice, and score questions over text, images, or JSON input
- Returns probabilities, confidence, selections, or weighted scores instead of free-form output
- Supports up to 400 questions per request and a dedicated /v1/decisions endpoint
- Provides prompt caching, GPT-Live integration, and regional processing for eligible customers
Languages
Known limitations
- API/developer product for bounded decisions, not a general-purpose chat, writing, or code-generation assistant
- Current Decisions pricing charges input tokens while output tokens are not charged; prompt-cache and long-context rates differ
- The current documented model is gpt-6-luna, so availability and behavior depend on that model and endpoint rollout
- Regional processing, Zero Data Retention, and HIPAA use require eligible customers, regions, and agreements








