OpenAI has officially released its Decisions API in public beta, marking a significant step towards more efficient and structured interactions with large language models. This new API is engineered to convert diverse inputs, including text and images, into definitive, typed answers that software applications can readily interpret and act upon.
According to OpenAI, the Decisions API processes requests approximately ten times faster than their existing Responses API. It specifically targets a common development challenge: prompting an LLM and then manually parsing its free-form text output into a usable label or decision. The new API seeks to automate and speed up this critical step.
Understanding the Decisions API
The Decisions API functions as a dedicated OpenAI endpoint for evaluating various inputs without generating extensive prose. Instead, it focuses on delivering precise, typed responses. Each request consists of three primary fields: the chosen model, the input data, and a series of specific questions. Currently, the sole supported model for this service is gpt-6-luna, though broader availability is anticipated in the coming weeks.
Requests are structured to include shared contextual evidence alongside a list of questions, each with a unique identifier, a defined type, and clear instructions. The API's response then provides an array of answers, indexed by the names provided in the initial questions.
Key Question Types
The Decisions API offers three distinct question types to facilitate diverse analytical needs:
- Predicate: This type assesses a given condition, returning a probability score between 0 and 1. An example use case could involve determining the likelihood of a product image displaying a defect like a crack or dent.
- Choice: Designed to select a single value from a predefined set of options, this type also delivers probabilities for each potential choice and a confidence metric for the selected answer.
- Score: The score type rates input against a sequence of ordered levels, indexed from zero. It calculates a probability-weighted average of these level indices, providing a nuanced assessment. OpenAI illustrates this with a severity example, where level probabilities of 0.1, 0.7, and 0.2 result in a score of 1.1, indicating a state between 'Workaround available' and 'Fully blocked'.
Distinction from Other OpenAI Tools
OpenAI clarifies the appropriate use cases for its various tools. The Decisions API is best suited for scenarios requiring probabilities, choices, or scores. In contrast, Structured Outputs should be utilized when the goal is to populate a custom JSON schema or generate detailed explanations. Function calling remains the preferred method when a model needs to invoke an external tool with specific arguments.
Performance and Pricing Details
OpenAI claims the Decisions API offers a substantial speed improvement, with responses approximately ten times faster than its Responses API. Early benchmarks from DevDay coverage suggested a decision time near 150 milliseconds, significantly quicker than the roughly 1.6 seconds for standard Luna calls. However, accuracy or calibration data for this endpoint has not yet been publicly released, with documentation recommending that developers establish their own thresholds using application-specific labeled examples.
Regarding cost, using the gpt-6-luna model incurs a charge of $0.10 per one million input tokens. Notably, there are no additional fees for output tokens, cache reads, or cache writes. Standard regional processing premiums and long-context multipliers, such as double input rates for prompts exceeding 272,000 tokens with Luna, still apply.
The API is accessible via a POST request to /v1/decisions on the OpenAI API platform and through a dedicated Playground environment. It supports various SDKs, including Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0, and Java 4.78.0. For eligible customers, compliance options include Zero Data Retention (ZDR) and HIPAA, with data residency available in the United States and Europe (EEA plus Switzerland). The Decisions API can also integrate with the Live API for voice-driven actions through client delegation.
Competitive Landscape: TypeSafe Jev
A notable competitor in this space is TypeSafe Jev, launched on September 15, 2026. Jev, a System One model, similarly provides typed values with calibrated probabilities. TypeSafe prices its input at $0.042 per one million tokens, with no charge for output, making OpenAI's base rate roughly 2.4 times higher. TypeSafe reports end-to-end response times between 70 and 500 milliseconds, measured from the US West Coast. While Jev supports up to 255 choices, it is currently in early access. OpenAI's Decisions API gains an advantage through its support for image input, extensive compliance options, and its status as an open beta release.
The OpenAI Decisions API represents a focused approach to extracting actionable, structured information from complex data inputs, potentially streamlining a wide array of AI-powered applications. Its current beta status means further development and independent evaluation will be crucial for its long-term adoption.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
