TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
OpenAI has opened the Decisions API in public beta for applications that need typed evaluations of text, images or both. It offers predicate, choice and score answers through a dedicated endpoint; OpenAI says it returns answers about 10 times faster than the Responses API. The service currently supports only gpt-6-luna, and general availability is expected in the coming weeks, according to OpenAI.
OpenAI has opened its Decisions API in public beta, introducing a dedicated endpoint for applications that need a model to classify, route or score text and images. The company says the endpoint returns typed answers about 10 times faster than the Responses API; the beta currently supports only gpt-6-luna.
The endpoint, POST /v1/decisions, accepts shared input and one or more questions, then returns an answers array. Each question has a unique name and a specified type, allowing an application to request a probability for a condition, a selection from a fixed list, or a rating against ordered levels. OpenAI says developers can try the API in its Playground before integrating it into an application.
The three question types serve different purposes. A predicate estimates the probability that a condition is true; a choice selects among supplied categories and returns probabilities for those options; and a score rates input against ordered levels, producing a probability-weighted average of their indices. OpenAI’s examples include checking a product image for visible damage, sending a billing complaint to the right department, and rating the severity of a software issue.
For image requests, the API accepts image content as inline base64 data URLs, combined with text in a user message. OpenAI’s documentation says hosted HTTP or HTTPS image links and file_id inputs are not supported by this endpoint. Developers also need supported SDK versions or later: Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0 or Java 4.78.0.
A Dedicated Route for Evaluations
The beta gives developers a separate interface for tasks where an application needs a bounded, typed answer rather than a generated response. A team could use a probability to flag material for human review, route a request according to a list of departments, or assign an ordered severity score. Because the question and answer format are defined in advance, applications can consume results without treating them as free-form prose.
OpenAI’s speed comparison may matter for applications that run many evaluations or need quick routing, but it is the company’s stated comparison, not an independently reported benchmark in the supplied documentation. The source does not provide test conditions, latency figures, or details about the baseline used for the “about 10 times faster” claim. Developers will need to test performance and answer quality with their own inputs before relying on it in production workflows.
The endpoint also makes a distinction between evaluation and broader response generation. OpenAI recommends using Structured Outputs with the Responses API when an application needs an object matching its own JSON schema, such as extracted fields or an explanation, and function calling when the model should request a tool call with arguments. That division helps developers choose an interface based on the kind of output their application needs.
Products Worth Considering
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
Logitech MX Brio Ultra HD 4K Webcam, Graphite
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Endpoint Frames Questions
The Decisions API is built around a request containing a model, shared input and questions. A single input can include text or messages with both text and images, while each question specifies the evaluation type and its instructions. The response echoes question names in its answers, giving application code a way to match results to requests.
For categorical decisions, developers supply the allowed values and descriptions. OpenAI advises using distinct options and including a fallback such as “other” when the categories do not cover every case; that result can be sent to a general review queue. For ordered ratings, the levels should be arranged from lowest to highest and describe the criteria for each level. These are configuration choices developers must define for their use case.
OpenAI describes the API as a tool for classification, routing and prioritization, rather than a universal replacement for its other API patterns. The public-beta status means the interface is available to developers, while the company still describes general availability as a future step. Only gpt-6-luna is listed as supported at launch.
Beta Limits and Performance Evidence
The supplied announcement does not specify pricing, rate limits, service-level commitments, regional availability, or whether the beta has restrictions beyond the listed model and image-input requirements. It also does not say whether additional models will be supported before or after general availability. These details should not be assumed from the endpoint examples.
The “about 10 times faster” comparison is attributed to OpenAI, but the documentation provided here does not explain the test setup, workload, or measurement window. No independent performance results are included. The example probabilities and confidence values are illustrative outputs, not reported accuracy results or guarantees for other inputs.
It is also unclear from the supplied material whether the endpoint’s behavior or request format may change during the beta. Developers evaluating it for consequential decisions will need to set their own thresholds and review paths; OpenAI’s examples show probabilities and fallback categories, but do not establish how a particular application should act on them.
General Availability in Coming Weeks
OpenAI says it expects to move the Decisions API from public beta to general availability in the coming weeks. The company has not provided a specific date in the source material. Developers can currently experiment in the Playground, then use the dedicated endpoint and supported SDK versions to test their own text or image evaluation tasks.
For teams considering deployment, the next practical step is to validate the endpoint against representative inputs, decide how probability and confidence values will be used, and define a fallback or human-review route for uncertain cases. Further details about timing, supported models, performance testing and beta terms remain dependent on information OpenAI has not provided in the supplied announcement.
Key Questions
What does the Decisions API do?
It evaluates text, images or both and returns typed answers: a probability that a condition is true, a choice from supplied values, or a score against ordered levels.
Which model can developers use?
The public-beta documentation lists gpt-6-luna as the only supported model at this time.
How are image inputs sent?
Images must be provided as inline base64 data URLs in a user message. The endpoint does not support hosted HTTP or HTTPS image URLs or file_id inputs, according to OpenAI’s documentation.
How does this differ from the Responses API?
Decisions is for predefined predicate, choice and score answers. OpenAI points developers to the Responses API with Structured Outputs for objects following a custom JSON schema, and to function calling when the model needs to request a tool call.
When will the API leave beta?
OpenAI says it expects general availability in the coming weeks, but the supplied announcement gives no specific launch date.
Source: hn
