OpenAI has launched the Decisions API as a public beta. It runs only on GPT-6 Luna and returns not running text but typed answers: probabilities, choices and scores. Only the input is paid for, at $0.10 per million tokens; the output costs nothing.

What happened

According to most sources the public beta began on 6 October 2026. One source, by contrast, names 29 September, DevDay, as the start of a limited preview. The interface is reachable at POST /v1/decisions and currently supports only the model gpt-6-luna. OpenAI says general availability is coming "in the coming weeks", without a date.

How the interface works

The Decisions API evaluates text and images and answers questions about them in fixed forms. Three kinds are described:

  • Condition: the probability that a condition holds.
  • Choice: the pick from given options, with the options' probabilities and a confidence.
  • Score: a rating on an ordered rubric, with the probabilities of the levels and a confidence.

In all three cases the result is a value in a fixed form that a program can process further.

A request consists of three parts: the model field, which currently has to be gpt-6-luna, an input field with the shared material, and a list questions. The response contains a list answers; each question carries a unique name that the API returns with the answer. For a choice, the API returns the chosen value, the options' probabilities and a confidence. For scores it returns the probability-weighted average of the levels, the probabilities of the levels and the confidence. One guide advises treating confidence separately from the options' probabilities and accounting for refusals per question.

Cost and speed

The rate for input is $0.10 per million tokens; output is not charged. Surcharges for processing in certain regions and for very long inputs still apply. Other requests to gpt-6-luna that do not run through /v1/decisions are billed at the usual rate.

On speed, OpenAI says the interface answers about ten times faster than the Responses API. One source gives about 150 milliseconds. The sources evaluated found no independent measurement.

Background

GPT-6 Luna belongs to OpenAI's GPT-6 line. According to reports, users on the Free and Go plans in ChatGPT get GPT-6 Luna, and in the API GPT-6 Sol and Luna have been available since 22 September 2026. The Decisions API is thus an additional way into this model, tailored to a single kind of task: delivering decisions instead of text.

On data protection and compliance, a comparison table reports that the beta supports Zero Data Retention, is suitable for HIPAA and offers data storage in the US and the EU. This statement was not found in an OpenAI document that could be read; anyone who depends on it should have OpenAI confirm it.

The sources also disagree on access. A guide on Hugging Face writes that POST /v1/decisions still answered with error 403 at the beginning of October with a normal API key, a block for certain accounts. Other sources describe the beta as open. Anyone who wants to use it should check access for their own account before building on it.

A note on interpretation: one source gives response times of 70 to 500 milliseconds; they belong to another product and not to OpenAI. The tenfold speed is a statement by OpenAI itself.

More on this topic:

What this means

For developers, the Decisions API is meant for cases in which an application needs a clear choice or a number, not running text. That can be derived from the answer types described: probability, chosen option, score level. Since only the input is charged, the bill depends on the length of the material, not on the length or number of the answers. It is important to handle confidence and refusals in your own code.

For companies that process data, the statements on Zero Data Retention, HIPAA and data storage in the US and the EU are important but so far documented only in a comparison table. They should secure the point directly with OpenAI before use.

For everyone else, the announcement is news from the engine room: it concerns the programming interface (API) for developers; what end users will notice is not stated in the sources.

A note on interpretation: the offering is a beta, and several statements come second-hand. The documentation and billing page of OpenAI are authoritative.

Fre­quently asked questions

What does the Decisions API return?

Typed answers instead of free text: the probability that a condition holds, a choice from given options, or a score against an ordered rubric. It evaluates text and images.

What does the Decisions API cost?

$0.10 per million input tokens; output tokens are not charged. Surcharges for regional processing and for very long inputs may apply. Other requests to gpt-6-luna outside /v1/decisions are billed at the usual rate.

How far along is the beta?

According to most sources the public beta began on 6 October 2026; one source names a limited preview at DevDay on 29 September. OpenAI says general availability is coming “in the coming weeks”, without a date.

Is it really ten times faster?

That is OpenAI's own statement compared with the Responses API. One source gives about 150 milliseconds. The sources evaluated found no independent measurements.