OpenAI Decisions API launches publicly with support for up to 400 questions

OpenAI made its Decisions API generally available on 22 October 2026. Designed for tasks that require a limited set of possible answers, the service is powered by GPT-6 Luna and lets developers get probability, choice or scoring responses from text, images and structured data. The API is available through a dedicated Decisions endpoint.

What kinds of decisions can the API make?

With the Decisions API, a developer first defines the question and the accepted answers, then provides the context needed to make a decision. OpenAI says the format is designed for applications that need to arrive at a specific result, rather than generate a long, open-ended response at every step.

The API offers three response types:

• Predicate: Estimates the probability that a condition is true, on a scale from 0 to 1.

• Choice: Selects one of a set of predefined options, and returns probabilities and confidence information alongside the answer.

• Score: Evaluates an input against ordered criteria levels.

OpenAI says the Decisions API can return results up to 10 times faster than GPT-6 Luna through the Responses API on certain tasks. This is a performance claim about the company’s own product; actual speed may vary depending on the workload and how the application is used.

Input limits and pricing

The generally available version supports native JSON inputs, images in question instructions and up to 400 questions per request. The company says this is twice the previous limit. It also offers prompt caching for repeated context and native integration with GPT-Live.

OpenAI’s published price is $0.10 per million input tokens. Cached input tokens cost $0.01 per million, and output tokens are not charged separately. Details on 30-minute and 12-hour cached writes are available in a separate pricing guide.

Regional processing is available in the United States and the European Union. Zero Data Retention is available only to eligible customers. As a result, where data is processed and how long it is retained are among the points teams evaluating the service should review separately.

Voice applications and real-world examples

The GPT-Live integration enables decisions to be made from a limited set of options during a voice conversation. In OpenAI’s example, an avatar’s facial reaction is selected based on the latest phrase in the conversation transcript. An application can use this output to determine what happens next.

Customer examples shared by the company include Box classifying business documents and routing them to the appropriate next step, Datadog assessing risk in agent commands, and Ramp attributing AI costs to jobs. In its own test, Speak reports reducing the time between actions from 1.4 seconds to 0.13 seconds using the GPT-Live integration. These are customers’ own measurements and do not mean every application will achieve the same results.

What it could mean for architecture and visualization workflows

For architecture firms and visualization teams, the API could be considered for automating repetitive steps with predefined response options. Potential use cases include classifying image or text inputs against criteria set by a team in its own systems, or routing a request to the appropriate predefined option. These are not ready-made architecture features; they are possible applications based on the input and response types the API supports.

When evaluating the service, teams should consider integration with existing software, the region where images and project information will be processed, data retention terms and input volume. Although pricing is stated only for input tokens, total cost depends on usage and whether caching is used. OpenAI recommends trying the Decisions Playground with your own text or images for initial testing.