OpenAI released its Decisions API in public beta on October 6. It is for application decisions such as sorting support requests or flagging product photos, not for generating a chat response. Send evidence and named questions to POST /v1/decisions with gpt-6-luna; the response is a typed answer rather than prose.

  • Three question types: predicate estimates whether a condition is true; choice picks from supplied categories and returns their probabilities; score averages probabilities over ordered rubric levels. Independent questions can share one input in a request.
  • Getting started: Try questions in the Playground, then call the endpoint with an OpenAI API key, model, input, and questions. Inputs can be text or text plus images; image data must be supplied as inline base64, not a hosted URL or file ID.
  • Cost: OpenAI lists $0.10 per million input tokens for this endpoint, with no output-token charge; regional and long-context premiums can apply. The docs claim roughly 10× faster answers than the Responses API, but that is OpenAI’s comparison, not an independent benchmark.

Limit: This is a classifier, not a calibrated guarantee or a replacement for an explanation. Only gpt-6-luna is supported in beta. OpenAI recommends testing thresholds on labeled examples and routing uncertain or costly decisions to review. Use Responses API Structured Outputs if you need extracted fields or narrative reasoning instead.

In the HN discussion, Simon Willison reports a direct call to the endpoint and shares its returned probabilities; that is an individual test, not evidence of accuracy across workloads. This capture is based on the official documentation and discussion, not our own hands-on test.