OpenAI Opens Decisions API Public Beta, Returning Typed Answers on GPT-6 Luna

The service targets a pattern developers already work with: prompt a language model, then parse its text into a label. A request carries three fields — model, input and questions. Each question has a unique name, a type and instructions, and the response comes back as an answers array keyed by those names. The only supported model today is gpt-6-luna, whose model card lists a 1,050,000-token context window; its parameter count is not disclosed. The API is available on OpenAI-hosted infrastructure only, through POST /v1/decisions, with no open weights and no self-hosting.

Three question types are supported. A predicate checks a condition and returns a probability from 0 to 1, for example whether a product photo shows a crack, tear or dent. A choice picks one value from options supplied by the caller and also returns per-option probabilities and a confidence field. A score rates input against ordered levels indexed from 0, calculated as a probability-weighted average of the level indices. OpenAI's severity example makes the arithmetic concrete: level probabilities of 0.1, 0.7 and 0.2 yield a score of 1.1, a value between the levels labeled "Workaround available" and "Fully blocked."

OpenAI draws a line between this endpoint and its existing tools. Decisions is for probabilities, choices or scores; Structured Outputs is for filling a caller's own JSON schema or writing explanations; function calling is for cases where a model must request a tool call with arguments.

On speed, OpenAI's documentation claims roughly 10x faster responses than the Responses API, and DevDay coverage put a decision near 150 milliseconds against about 1.6 seconds for regular Luna calls. OpenAI has not published accuracy or calibration data for the endpoint, and the documentation advises setting thresholds using labeled examples from the caller's own application.

With gpt-6-luna, input costs $0.10 per 1 million tokens, with no output-token, cache-read or cache-write charges. Regional processing premiums and long-context multipliers still apply, and Luna's model card prices prompts above 272K tokens at 2x input rates. The endpoint ships with a Playground in addition to the API, and OpenAI lists minimum SDK versions of Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0 and Java 4.78.0. Zero Data Retention and HIPAA are available to eligible customers, with data residency in the United States and Europe, meaning the EEA plus Switzerland. Decisions can also drive actions through client delegation with the Live API. OpenAI expects general availability in the coming weeks.

The closest rival is TypeSafe Jev, a System One model that returns typed values with calibrated probabilities and has been in early access since September 15, 2026. TypeSafe prices input at $0.042 per 1 million tokens with free output, which puts OpenAI's base input rate at about 2.4x higher. TypeSafe reports 70 to 500 milliseconds end-to-end, measured from the US West Coast, and Jev supports up to 255 choices for a single question but remains in early access. Jev accepts text and structured state; image input, compliance options and open beta access are where OpenAI's offering differs.

Editor's Summary

OpenAI has put its Decisions API into public beta, letting developers obtain probabilities, choices and scores instead of prose from gpt-6-luna at $0.10 per 1 million input tokens. The company claims roughly 10x faster responses than its Responses API, though no independent accuracy or calibration data has been released. TypeSafe's Jev undercuts it on input price but remains in early access and does not accept images.