OpenAI puts Decisions API in public beta

OpenAI says the Decisions API is in public beta and that it expects general availability in the coming weeks.[2] Unite.AI reports the release date as October 6, 2026.[1] AlphaSignal says the public beta is available to all developers.[3] OpenAI describes the API as a way to help applications classify content, route requests and prioritize work.[2] That makes the beta relevant to teams evaluating model-based steps in existing workflows, while its current availability and feature limits still matter when planning a production rollout.

Developers can ask for a predicate probability, a selection from a fixed set of values, or a rubric-based score.[2] A predicate returns an estimated probability from 0 to 1 that a condition is true; a choice selects one of the values supplied by the developer.[2] A score is a probability-weighted average of the level indices, and choice and score questions return probabilities over discrete options.[2] Developers call the dedicated `POST /v1/decisions` endpoint, and OpenAI offers a Playground to try the API before writing code.[2]

OpenAI says the API returns typed answers about 10 times faster than its Responses API.[2] AlphaSignal reports that a launch demo routed 10,000 customer requests among Billing, Technical and Sales, and showed 150 milliseconds per request for Decisions versus 1.6 seconds for Responses.[3] AlphaSignal describes that difference as roughly 10.7 times.[3] It also says OpenAI did not publish enough methodology to reproduce the benchmark, including latency percentiles and payload sizes.[3] Teams should treat the demo as a reported result, not a complete basis for forecasting performance in their own workloads.

The only model currently available for the Decisions API is `gpt-6-luna`.[2] Image inputs must be inline base64 data URLs; hosted HTTP or HTTPS image URLs and `file_id` inputs are unsupported.[2] Pricing, tuning support and candidate-answer limits remained unpublished at beta launch, according to AlphaSignal.[3] Developers who need an object conforming to their own JSON schema can use Structured Outputs with the Responses API, OpenAI says.[2] These details can shape a pilot: verify that the input format and response type fit the workflow, and assess the beta’s open commercial and technical questions before committing to broader use.