OpenAI launches Decisions API for faster bounded decisions in agents
OpenAI announced the general availability of the Decisions API on October 22, introducing a specialized endpoint that produces bounded answers an application can turn directly into an action. Instead of generating free-form prose, the service returns a probability, selects a predefined option, or calculates a score from an ordered scale. It is designed for tasks such as classifying inputs, routing requests, and choosing an agent's next step. [1]
Three answer types
The endpoint uses GPT‑6 Luna and accepts context as text, images, or JSON. A “predicate” estimates a probability from zero to one that a condition is true; “choice” selects among developer-defined values and also returns probabilities and confidence; “score” computes a probability-weighted average across ordered rubric levels. One request can contain as many as 400 questions about shared context, twice the previous limit. [1]
OpenAI describes the API as a path to near-real-time decisions and says it can respond up to ten times faster than GPT‑6 Luna through the Responses API. That figure is a provider measurement, not an independent evaluation. It also does not establish the same advantage across every region, workload, or input class. Unite.AI confirmed the release and described how the endpoint works, although its report relies primarily on OpenAI's announcement and documentation. [2]
Changes since beta
The generally available version adds native GPT‑Live integration for decisions during voice conversations, prompt caching when requests reuse the same context, and inference residency in the United States and Europe. Developers can also pass native JSON and place images in question instructions. OpenAI says it has added capacity in several Asian regions. [1]
The GPT‑Live integration targets situations in which a voice assistant must decide whether to ask for clarification, route a conversation, or escalate a case. Outside voice applications, the same structure can prioritize work, categorize transactions, or assess an image against a defined condition. According to Unite.AI, GPT‑6 Luna is currently the only supported model and requests use a dedicated endpoint. [2]
A control component, not a complete agent
The Decisions API is not an autonomous agent and does not replace application logic. Developers define the questions, allowed answers, and downstream behavior. That separation may reduce the ambiguity of free-form output and simplify validation, but a probability or selection remains a model inference. Sensitive workflows still need thresholds, permissions, logs, exception paths, and human review proportionate to the consequences of a wrong decision.
General availability moves a capability introduced in beta in early October into a stable platform product, but broad external evidence is still missing on accuracy, probability calibration, sustained latency, and behavior under distribution shifts. Teams considering adoption should compare it with conventional rules and specialized models on representative data rather than assuming that an aggregate speed claim determines quality.
The release also illustrates a more specialized direction for agent infrastructure: separating language generation from the small, frequent decisions that govern a workflow. If that specialization lowers latency and cost without reducing reliability, it may be useful for interactive products and high-volume operations. The relevant test will be improvement across the whole system—including errors, escalations, and oversight—not merely endpoint speed.
— Nova Rivera