OpenAI takes Decisions API out of beta, targeting faster app decisions

AI infrastructure · October 24, 2026

OpenAI made its Decisions API generally available on October 22, expanding a service designed to help software classify information and select its next action without generating a full conversational response. The release uses the existing GPT-6 Luna model and adds capabilities for production applications, rather than introducing a new foundation model. OpenAI’s announcement

The timing matters: OpenAI first announced Decisions in limited preview at DevDay on September 29, then opened its public beta on October 6. This week’s development is its move out of beta, alongside a larger set of supported inputs and operational features.

More context, more questions

The October 22 changelog confirms prompt caching, native JSON inputs and images within individual question instructions. A request can now contain up to 400 questions, double the beta limit. OpenAI also says the service now supports inference residency in the US and EU.

Voice integration is part of the expanded offering, but predates general availability: OpenAI’s changelog dates native GPT-Live integration to October 15. The voice documentation describes two routes: letting GPT-Live call Decisions directly, avoiding an extra round trip through an application, or triggering decisions as transcript updates arrive. That could help a voice interface respond to a changing request before a speaker finishes.

A smaller job than a chatbot response

Developers supply the question and the permitted form of answer. A predicate estimates whether a condition is true; a choice selects a predefined option; a score evaluates input against ordered criteria. The developer guide identifies GPT-6 Luna as the only supported model.

For a business application, the appeal is a faster intermediate step: deciding which queue should receive a document, for example, before another system handles it. The commercial significance lies in how often those small decisions recur. Reducing their overhead could improve an entire workflow, even when the system doing the final writing or complex reasoning stays unchanged.

Low headline pricing has qualifications

The Decisions pricing guide lists uncached input at $0.10 per million tokens, with no output-token charges. Cache reads use Luna’s standard rate of $0.01 per million tokens. The current pricing table lists cache writes at $0.125 per million tokens for 30-minute retention and $0.25 for 12-hour retention.

Those rates do not mean every request receives the cache discount. Reuse requires a matching shared-input prefix of at least 1,024 tokens, and question content is outside that reusable prefix. Regional-processing premiums and long-context input multipliers also apply. These details make workload shape important when estimating savings.

Speed claims still need application-level testing

OpenAI says Decisions can be up to 10 times faster than GPT-6 Luna through the Responses API. That is the company’s comparison, not a result independently reproduced by PromptsZone, and it should not be read as a guaranteed tenfold improvement in an entire application.

Nor does a structured answer eliminate uncertainty. OpenAI’s guidance on interpreting results recommends setting thresholds with labeled examples from the actual application, taking account of false positives and false negatives. Its examples also handle refused answers, so applications need a path for cases that cannot be classified.

The release makes narrow AI judgments easier to incorporate into everyday software. Whether that produces a better product depends on more than latency: teams still need evidence that their chosen questions, fallback behavior and review thresholds work on real traffic.

Reporting note: Based on OpenAI’s announcement, release history and documentation checked October 24, 2026. PromptsZone has not independently tested the API; implications above are analysis.

Archival photo: Coolcaesar, “1515 Third Street,” via Wikimedia Commons, June 15, 2025, under CC BY 4.0. The building housed OpenAI’s headquarters when photographed. Display resized; not a photograph of this release. No photographer endorsement implied.