Key Notes
- OpenAI has made the GPT-6 Luna-powered Decisions API generally available for predefined choices, probabilities and scores.
- The company claims up to ten times faster decisions than Luna through the Responses API, with support for images, JSON and four hundred questions per request.
- Input starts at ten cents per million tokens with no output charge, while cache writes and processing options can affect the final bill.
OpenAI has made its Decisions API generally available, giving developers a dedicated way to ask AI a question and receive a predefined choice, probability or score. Built on GPT-6 Luna, the service targets the small judgments that repeatedly interrupt an application’s workflow: where a request should go, how urgent it is, or what an agent should do next.
The company’s October 22 announcement claims performance up to 10x faster than using the same model through the Responses API for these bounded tasks. That is a comparison for focused decision-making, rather than a claim that Luna has become ten times faster at every kind of AI work.
A Decision an Application Can Use Immediately
Developers supply context and define what the system may return. The developer guide describes three question types: a predicate estimates whether a condition is true, a choice selects an allowed option, and a score evaluates an input against ordered levels. Choice and score results also include probability information and a separate confidence field.
Consider an equipment supplier handling a customer’s message and a photograph. Its application could check whether damage is visible, classify the request and assess urgency. The result can feed directly into routing rules, while an uncertain case goes to a reviewer. This is an illustrative workflow; the developer must still determine which actions a model’s answer is allowed to trigger.
The API accepts text, images and native JSON, with support for up to 400 questions per request, double the beta limit. Questions that examine the same evidence can share a request. A question that depends on an earlier answer needs a subsequent request, so the allowance does not turn one call into 400 sequential reasoning steps.
Voice Assistants Can Make Choices Mid-Conversation
The release adds native integration with GPT-Live. A support assistant could use a decision to choose between asking another question and escalating to a person; an animated character could select a facial expression from the conversation. The voice system handles the exchange while the decision component supplies a focused judgment.
For a user, the useful improvement would be a shorter pause before the service responds or changes course. It is still the complete application that determines the experience: transcription, network traffic, downstream tools and the chosen review process can all affect waiting time. A faster model call alone does not guarantee a faster resolution.
Ten-Cent Input Pricing, With Caching Details to Watch
Uncached input costs $0.10 per million tokens, with cached input at $0.01 per million and no charge for output tokens. At the uncached base rate, one billion input tokens would cost $100. That is a token-cost calculation, not the price of running a complete customer-service operation.
OpenAI’s current pricing guide lists separate cache-write rates, and the Decisions documentation says regional processing premiums and long-context multipliers apply. Reusing a qualifying context prefix can reduce input costs, but filling a cache is not the same billing event as reading from it.
The launch reinforces a category we have followed through Microsoft’s Decision-1 and Cloudflare’s Clef models. These products focus on selecting among allowed outcomes, making latency, cost and decision accuracy more relevant comparisons than the quality of a long generated answer.
The important test remains whether the system selects the right outcome on real inputs. Teams need labeled examples, thresholds for escalation and a way to handle unfamiliar cases. A probability estimate can help organize that process; it cannot replace checking what happens when the software acts on the answer.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
