OpenAI announced the full launch of its Decisions API, powered by GPT-6 Luna, on October 22, local time. In terms of pricing, the Decisions API charges US$0.10 per million tokens for input (approximately HK$1), while cached input costs US$0.01 per million tokens (approximately HK$0). The entire OpenAI team is working to integrate the Decisions API into products and business workflows. The core design of the OpenAI Decisions API is to focus GPT-6 Luna's intelligence on questions with limited answers rather than generating narrative text.
Developers define a question and its allowed answers, then provide context in text, image, or native JSON format. The API returns one of three types of answers: predicate (the probability that the condition is true, ranging from 0 to 1), choice (selecting one from predefined options, along with probability and confidence), or score (probability-weighted scoring of input on an ordered scale).
This "narrow interface" design has a clear intent: classification, routing, scoring, and prioritization are small decisions frequently executed by applications and agents within workflows. While a single call may seem simple, delays and costs can quickly accumulate at scale. OpenAI officials stated that the Decisions API responds ten times faster than calling GPT-6 Luna through the Responses API. The pricing structure continues the logic of the Decisions API's "narrow interface."
Input is charged at US$0.10 per million tokens (approximately RMB 0.67), while hits to the cache are charged at US$0.01 per million tokens (approximately RMB 0.067). The API charges only for input tokens, with output tokens being free. Since the Decisions API returns typed answers rather than generated text, there is no cost associated with the output itself. The pricing for cached input further reduces the cost of individual decisions in scenarios where context is reused.
OpenAI Decisions API Provides More Efficient Decision Support
The Decisions API entered public testing on October 6, 2026, at which point a single request supported up to 200 questions. Upon full launch on October 22, the question limit per request was increased to 400, doubling the capacity from the public testing phase. Additionally, since public testing, native GPT-Live integration, prompt caching for reused context, and reasoning data residency support in the U.S. and EU have been added. Developers now support incoming native JSON input and the inclusion of images in question prompts.
OpenAI also stated that it has continuously optimized latency performance, including expanding service capacity in several regions across Asia.
The applicability of the OpenAI Decisions API has been validated internally at OpenAI. According to the official OpenAI blog, the internal dots team used the Decisions API to reduce response times and enhance interactivity; the personal finance team utilized it to classify transaction descriptions; and the communications team employed it to categorize reports by topic and scoring, determining whether they reinforce or challenge key narratives, thereby identifying content that requires further review. Client applications are also expanding.
OpenAI mentioned that clients are using the Decisions API for various scenarios, from attributing AI costs to making language learning experiences more interactive. In the GPT-Live integration scenario, customer service agents can quickly determine whether to ask for clarification, route requests, or escalate issues; virtual avatars can trigger facial reactions based on user dialogue, enabling more natural real-time interactions.
The Decisions API is not a replacement for the Responses API but rather a segmentation of the latter's use cases. If text generation, custom JSON schema, tool invocation, streaming output, or maintaining conversational state is required, the Responses API remains the correct choice; however, when tasks involve classification, routing, scoring, or gating, and require probabilities and confidence, the Decisions API offers a dedicated path with lower latency and a cost structure based solely on input.
This division of labor also reflects a trend in AI application architecture: separating deterministic, high-frequency judgment tasks from general generation interfaces and delegating them to narrower, faster, and cheaper dedicated endpoints.