OpenAI launches Decisions API for 10x faster classification

OpenAI has released the Decisions API, a new endpoint designed to evaluate text, images, or both and return typed answers significantly faster than its standard Responses API. According to the official developer documentation, the new service is approximately 10x faster for these specific tasks, allowing developers to classify content, route requests, and prioritize work with lower latency. The API is currently in public beta, with a general availability (GA) release expected within the coming weeks.

The primary use case for the Decisions API is when an application requires a specific type of structured answer rather than free-form generation. Developers can request the probability that a condition is true, select a choice from a fixed set of options, or calculate a score against a defined rubric. This distinguishes it from the Responses API, which is better suited for generating objects that follow custom JSON schemas or for function calling where the model needs to request tool arguments.

Technical Specifications and Availability

The Decisions API is currently accessible only via the gpt-6-luna model. Developers must use the dedicated POST /v1/decisions endpoint to interact with the service. To utilize the SDK examples provided in the documentation, specific minimum versions of the OpenAI SDKs are required: Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0, and Java 4.78.0.

A request to the Decisions API consists of three main parts, and the response includes an answers array. Each question in the request must be assigned a unique name, which the API echoes back in the response to identify the corresponding answer. This naming convention allows developers to map multiple questions in a single request to their respective results programmatically.

The API supports two primary types of discrete answers: choice and score. Both return probabilities over discrete options. The choice type is intended for categories without an inherent order, such as routing a ticket to specific departments. The score type is designed for ordered levels, such as severity ratings. It calculates a probability-weighted average of the numeric indices of the options, producing a score that can fall between defined levels. This allows for more granular prioritization than a simple binary or categorical output.

Image and Text Evaluation

One of the key features of the Decisions API is its ability to process multimodal inputs. Developers can combine input_text and input_image parts in a user message to evaluate images alongside instructions or other context. For example, the documentation illustrates a use case where a predicate question checks a product photo for visible damage, such as cracks, tears, or dents. The API returns a probability representing the model’s estimate that the condition is true, which developers can use to flag photos for human review based on a chosen threshold.

However, there are specific constraints on image handling. Images must be provided as inline base64 data URLs. Hosted HTTP or HTTPS image URLs and file_id inputs are not supported by this endpoint. This requirement means that developers must handle the encoding of images on the client side before sending the request, which may impact payload sizes and network performance compared to passing a simple URL.

Developer Workflow and Integration

For developers, the Decisions API offers a streamlined path for applications that rely on high-volume classification tasks. Instead of using the more general-purpose Responses API, which may incur higher latency and cost for simple routing or scoring tasks, the Decisions API provides a specialized interface. The documentation suggests using the Decisions API when the application needs one of the specific answer types mentioned (probability, choice, or score). Conversely, Structured Outputs with the Responses API should be used when generating complex objects with custom JSON schemas, such as extracted fields or written explanations.

The API is available for experimentation in the OpenAI Playground, allowing developers to test questions and inputs before writing code. This is particularly useful for tuning the prompts and rubrics to ensure the model’s probabilities align with business logic. For instance, in a team routing scenario, a choice question can select a department from options like "Driving" or "Team routing," while a score question can assess the severity of a support ticket.

What to Watch

• General Availability Timeline: The API is in public beta, and OpenAI expects GA in the coming weeks. Developers should monitor the documentation for any changes in model availability or endpoint stability before relying on it for production workloads.
• Model Limitations: Currently, gpt-6-luna is the only model available for the Decisions API. As other models are added, the performance and accuracy of the typed answers may vary, requiring re-tuning of prompts and thresholds.
• Image Input Constraints: The requirement for inline base64 data URLs may pose challenges for applications dealing with large images or high-throughput scenarios. Developers should evaluate the impact on network latency and payload size.
• SDK Version Requirements: Ensure that all client applications are updated to the minimum required SDK versions to access the new endpoint and features. Older versions will not support the POST /v1/decisions endpoint.