OpenAI Decisions API Goes Live for Fast Agent Calls

OpenAI has moved one of its most practical developer tools out of beta. According to Unite.AI, the OpenAI Decisions API became generally available on October 22, 2026, powered by the GPT-6 Luna model and priced at a fraction of the cost of a full chat-style model call.

The idea is simple, and it speaks directly to how agents actually work. Most production agents do not only generate long answers. They constantly make small, bounded calls: is this message spam, which department should handle this ticket, how severe is this bug. The OpenAI Decisions API is built for exactly those calls, returning a typed answer instead of an essay.

Developers can ask three kinds of question. A predicate returns a probability that a condition is true, such as whether a photo shows visible damage. A choice picks one option from a fixed set, with a probability for every option and a confidence figure. A score rates an input against a rubric, returning a probability-weighted average across ordered levels.

That structure matters because it removes guesswork from the other end of the call. An application does not have to parse free text to find the answer, and it does not have to wonder whether the model answered in the expected format. The answer arrives in a shape the code can use immediately, which is why the tool fits so naturally into agent workflows.

Speed is the other headline. OpenAI says the endpoint answers up to ten times faster than asking the same model through its standard Responses API, with response times reported around one to two hundred milliseconds. According to Vector Wire, OpenAI's API product lead said the service shrank response times from one to three seconds down to roughly that range, and early trial customers include Box, Datadog, Ramp and Speak.

What changed since the beta

The general availability release adds several features that were missing from the public beta. A single request can now carry up to four hundred questions, double the previous limit, so a batch of classifications can travel in one call. Inputs can now be text, images or native JSON, and images can appear inside individual question instructions so several images can be compared against a shared reference.

Voice applications gain native integration with the company's live voice service, so a spoken app can react as a person speaks rather than waiting for a full turn to finish. Prompt caching has also arrived, lowering costs when the same context is reused across calls. The OpenAI developer community announcement lists these changes alongside the pricing, and it notes that inference residency is available in the United States and Europe.

Pricing is deliberately simple. Uncached input costs ten cents per million tokens, cached input costs one cent per million tokens, and there are no output-token charges at all. That last point is unusual: most model calls charge for both directions, but a decision endpoint returns such a small answer that charging for output would add complexity without adding much revenue.

Why agents, not chatbots, are the real audience

The timing tells its own story. Agents are moving from demonstrations into production systems, and production systems live or die on thousands of small decisions. A support agent that classifies every incoming message, a coding agent that scores every test failure, or a research agent that routes every source to the right bucket will make far more of these calls than it will generate long reports.

That volume changes the economics. At the new price, a million short classifications cost a few dollars rather than a sum that needs a budget meeting, which means developers can afford to put a decision step wherever a human would once have made a judgment call. The result is agents that check their own work more often, rather than agents that simply run faster.

There are limits worth noting. The endpoint runs on one model only, and it is designed for answers, not generation. When a task needs extracted fields or a written explanation, OpenAI points developers back to structured outputs on its standard API. The Decisions API is a specialist tool, and its strength comes from refusing to be a general one.

Competition is already visible. The category of fast, structured decision models was pioneered by smaller specialists, and OpenAI's entry puts a frontier lab directly into that market. Independent benchmarks cited by analysts show the specialist models still leading on some decision benchmarks, which suggests the contest will be fought on accuracy and latency, not price alone.

For builders, the practical takeaway is straightforward. The OpenAI Decisions API gives agents a cheap, fast way to answer the small questions that pile up inside every workflow. Used well, it could make agents more reliable, because a system that can afford to check itself at every step is less likely to drift. Related coverage on this site includes reporting on recent model launches and on the price war reshaping enterprise AI contracts.