OpenAI Decisions API Brings Bounded Choices to General Availability

A bounded-decision endpoint enters general availability

OpenAI announced on October 22 that its Decisions API was generally available with GPT-6 Luna through a dedicated endpoint. It is intended for a narrow question whose answer comes from a defined set: classify an input, route a request or choose the next action in a workflow. The API accepts text, images and structured inputs. This is different from asking a general chat model to produce a long free-form response; a caller must define the candidate decisions and interpret the returned result in its own business context.

OpenAI claims responses can be up to ten times faster than GPT-6 Luna through the Responses API for the bounded tasks it describes. Customer examples report other ratios and one p95 below 280 milliseconds, each from a particular evaluation or deployment. These are supplier and customer claims, not a cross-vendor benchmark performed for this article. A ten-times headline cannot be assumed for every input length, region, cache state or decision set.

Cost and eligibility have conditions

The original product post quotes a base price of $0.10 per million input tokens and $0.01 per million cached input tokens, with no output-token charge for this endpoint. These are announcement-day rates, not a durable quote; teams must check the current pricing guide before budgeting. Regional processing in the United States and European Union is listed, while Zero Data Retention is available to eligible customers. Eligibility is a condition, not a default guarantee for every account.

A bounded decision can be useful in a hot path because a small latency difference repeats across thousands of workflow steps. But an inexpensive, fast classification can also propagate a mistake quickly. The appropriate comparison uses a labeled set from the actual application, measures accuracy and abstention alongside p50 and p95 latency, and includes retrieval, network and approval time. A customer testimonial reporting the same accuracy on 30 insurance cases does not establish general accuracy across industries.

Keep decisions within their authority

Consider a router that assigns incoming documents to teams. Its candidate labels, fallback behavior and confidence threshold should be specified before use. A security-sensitive action should still pass explicit policy checks and human review where required. A model result is evidence for a decision process, not authorization to execute a shell command or disclose a document. This distinction matters even when the API is generally available.

The release makes bounded model decisions a first-class API product. It does not say that all decision problems are solved, that every task fits Luna, or that the cited speed and price remain constant across future versions. The original announcement date is October 22; this October 26 article records the claims and a test plan for readers evaluating adoption.