OpenAI made its Decisions API generally available on October 22, moving a tool for small, bounded AI judgments beyond its October public beta. The current developer guide now spells out the request, answer types and pricing. For a developer deciding whether to put a model in an application’s routing path, the useful distinction is precise: Decisions supplies an answer to a question the application defines; the application still decides what to do with it.

The October 22 launch names additions since beta: native JSON input, images inside individual question instructions, prompt caching, native GPT-Live integration and up to 400 questions in one request, up from 200. OpenAI also says it added capacity in several Asian regions. Its API changelog records the beta release on October 6 and the GA update on October 22. Our earlier preview account described a product whose public contract and price were still unsettled; the published guide and reference now give builders a specific interface to assess.

What the API returns

The documented call is POST /v1/decisions with model, shared input and a questions array. The only currently supported model is gpt-6-luna. Shared input can be text or an ordered history of supported messages and tool calls. Each named question specifies its type and instructions, plus allowed choices or rubric levels where needed. The response’s answers array follows question order and echoes names so the application can match results to questions. These are descriptions of OpenAI’s guide and API reference, not a BIG CHANGE integration test.

A predicate returns an estimated probability from zero to one for a stated condition. A choice selects one developer supplied string or Boolean value and returns probabilities and a confidence field. A score applies ordered rubric levels and returns a probability weighted average of their numeric indices, which may fall between levels. A question can instead receive a refusal while other questions in the same request receive answers. That reference behavior matters to code that expects every answer to contain a choice or score.

OpenAI distinguishes these typed judgments from generating an arbitrary JSON object with Structured Outputs or asking a model to call a tool. The guide says tool calls and their outputs may be supplied as evidence, but the consuming application executes any selected action. A returned choice therefore does not grant permission to perform that action. Authorization, side effects and a fallback for refusals belong in the surrounding application.

What changed for inputs and control

The GA guide accepts input_json in a user message, so structured evidence need not first be turned into prose. Images can be shared evidence or included in a particular question’s instructions. Public HTTP or HTTPS image URLs and inline base64 data URLs are documented; file IDs are not. The reference limits a JSON content part to 256 KiB of compact UTF-8, 10,000 values and 32 container levels. Files, audio and item references are outside the documented Decisions input. GPT-Live is a separate native integration that can trigger Decisions from transcript events or delegate to it during a live session; it does not make raw audio a Decisions request input.

The 400-question ceiling is a batching limit, not a 400-step workflow. OpenAI says to group independent questions against the same input, and to send a later request when one decision depends on an earlier answer. Its guide recommends distinct choices, observable criteria and clearly separated rubric levels. That is also a useful test of fit: if a decision needs a free-form explanation or an open-ended tool plan, the three bounded answer types cannot express it directly.

Price, access and evidence to check

For Decisions using gpt-6-luna, OpenAI lists $0.10 per million uncached input tokens, $0.01 per million cached input tokens, and no output-token charge. The pricing guide lists $0.125 per million tokens for 30-minute cache writes and $0.25 for 12-hour writes at the standard rate. The Decisions guide says regional processing premiums and long-context input multipliers apply. A cache hit requires a matching shared-input prefix of at least 1,024 tokens; question content does not count toward that threshold. These are token rates, not a per-decision quote: input size, image tokens, cache behavior and the number of requests determine an application’s bill.

OpenAI says regional processing is supported in the US and Europe, while data residency is available in supported API regions. Zero Data Retention and HIPAA use have eligibility requirements. Developers need to check their project’s model access and rate limits, plus the data-controls terms, for the region and handling they require. The organization usage endpoint reports request counts and tokens; a request containing many questions counts as one model request, according to the guide.

OpenAI advertises answers up to 10 times faster than GPT-6 Luna through the Responses API for bounded questions. The launch gives that comparator, but does not publish the test set, sample size, latency statistic or request settings behind the figure. Its customer examples report separate workloads: Box, for instance, says route selection was about five times faster than Responses on 30 certificate-of-insurance cases in its evaluation. A useful evaluation would label examples from the actual routing task, compare false positives and false negatives at the chosen threshold, count refusals, and measure latency and token usage under the expected request mix. OpenAI itself recommends labeled examples and thresholds tied to error costs. BIG CHANGE has not run such an evaluation.

The big change

  • The October 22 GA provides a published, typed API contract for bounded questions, with one supported model and a 400-question request limit.
  • JSON, question-specific images, caching and GPT-Live integration widen where that contract can be used; the application remains responsible for execution and permissions.
  • The low input-token rate is a starting price. Accuracy, refusals, latency, cache hits and regional terms still need assessment in the intended workflow.