OpenAI makes its Decisions API generally available with 400 questions per request

What changed on 22 October

OpenAI's Decisions API is now generally available. The company's API changelog says the API has left public beta with gpt-6-luna as its model, and TechCrunch reported that it became available to all customers on Thursday, 22 October 2026. OpenAI first showed the product at its developer day on 29 September, according to TechCrunch, and opened it as a public beta in early October.

The GA release is more than a label change. OpenAI doubled the number of questions allowed in one request from 200 to 400, added prompt caching, JSON input, images inside individual questions, conversation history with tool calls, usage reporting and Private Link, and wired the API directly into its GPT-Live voice sessions.

What the Decisions API does

The Decisions guide describes a narrow job. Instead of generating free text, the API evaluates text, images and JSON and returns typed answers, which OpenAI says arrive about 10 times faster than through the Responses API. Requests go to POST /v1/decisions, and gpt-6-luna is the only supported model.

Each request has three parts: the model, a shared input that serves as evidence, and a list of questions. There are three question types. A predicate returns the probability, from 0 to 1, that a condition is true, such as whether a product photo shows a dent. A choice picks one value from a fixed set, such as which department should handle a ticket. A score rates the input against ordered levels, such as issue severity, and returns a probability-weighted average that can fall between levels.

Choice and score answers also include the full probability distribution and a separate confidence field. OpenAI tells developers to set their own thresholds using labeled examples from their application, choosing them according to the cost of false positives and false negatives. The API only answers; the application still runs whatever action the answer selects.

OpenAI is clear about where the product stops. If you need to extract fields into your own JSON schema, write an explanation or have a model call a tool with arguments, the guide points you back to Structured Outputs or function calling in the Responses API.

New since the public beta

According to the guide's list of changes since the beta, the generally available API adds:

Prompt caching, so shared input such as a long policy document or reference image can be reused across requests at a discount. By default the API places a cache breakpoint at the end of the shared input, and the reusable prefix must be at least 1,024 tokens. Question text is outside the cached prefix, so developers can keep the evidence fixed and vary the questions.

Images inside individual questions, so each question can carry its own reference picture to compare against the shared evidence, plus hosted image URLs in addition to inline base64 data.

An input_json content part, which lets developers pass structured data such as an order record directly rather than turning it into a string. OpenAI's example asks whether a refund request describes a duplicate charge that the order data actually shows.

Conversation history, so a decision can be made over a full exchange including function calls, custom tool calls and their outputs, along with organization usage reporting and Private Link connections.

Each request can now contain between 1 and 400 questions, and a request with many questions still counts as one request in usage reports.

Price and data controls

The pricing model is unusual. OpenAI's guide says uncached input to gpt-6-luna through /v1/decisions costs $0.10 per million tokens and that output tokens are not charged at all. Cache reads are discounted and cache writes are charged at gpt-6-luna's standard prompt-caching rates. Regional processing premiums and long-context multipliers still apply.

Because the answers are short typed values, removing output charges means the bill is driven almost entirely by how much evidence is sent. That rewards the pattern OpenAI recommends: put a large shared document or image in the cached input and ask many questions about it in one request.

For regulated users, the guide says the API supports Zero Data Retention and HIPAA use for eligible customers, data residency in all supported API regions, and regional processing in the United States and in Europe, covering the EEA and Switzerland.

Voice: decisions while the user is still talking

The most distinctive GA feature is the link to GPT-Live, OpenAI's real-time voice model. A separate voice guide explains two modes. With native delegation, Live calls the Decisions API itself when it recognizes a request your app can handle, skipping a round trip through your server. With triggers, decisions fire as transcript fragments of the user's speech or of Live's reply arrive.

OpenAI's worked example is a car voice assistant. When a driver says they feel cold, one decision request answers three questions at once: raise or lower the cabin temperature, switch the windshield defogger, and whether the request needs a full Responses API call for things like navigation. The app performs the actions, reports the confirmed state back, and Live tells the driver what happened.

The pattern fits OpenAI's recent voice work, including GPT-Live 1.1 with new voices and better accents. Fast typed decisions let a voice app act on a request mid-sentence instead of waiting for a large model to finish reasoning.

Early customers and the latency claim

Nikunj Handa, OpenAI's API product lead, told TechCrunch the main gain is speed, "shrinking things from one to three seconds to 100, 120, 180 milliseconds." He said OpenAI did not train a new calibrated decision model; the team reconfigured its existing ChatGPT Luna model to speed up inference and lower cost, and may add specific training later.

TechCrunch reported that early trial customers included Box, Datadog, Ramp and Speak. Andrew Hsu, chief technology officer at the language-tutoring company Speak, said the API cut the time between actions from 1.4 seconds to 0.13 seconds. OpenAI also used it to make its Dots assistant more responsive, according to the report. These are customer and vendor figures for specific workloads, not general benchmarks.

Handa also described where OpenAI wants to go next: if latency reached about 60 milliseconds, he suggested, developers could imagine calling the API on every frame of a video.

How it stacks up against dedicated decision models

OpenAI is entering a category it did not create. TechCrunch credits TypeSafe's Jev model with starting the market for fast decision models; AI Lookout covered TypeSafe's $870M Series A built on Jev. Since then, rivals have multiplied, including Liquid AI's D1 Ultra multimodal decision model and Base Labs' open mammoth-1 and elephant-1.

On public leaderboards, OpenAI's model is not the leader. TechCrunch reported that the AIM System One Index and the Artificial Analysis Decision Index tend to favor Liquid's D1 Ultra and Base Labs' Elephant-1, with TypeSafe's latest Jev seventh and third, and OpenAI's decision model eighth and fourteenth respectively. TypeSafe's models are currently the most queried on OpenRouter, the report added.

OpenAI's pitch is breadth rather than leaderboard position: image understanding from Luna, which Handa called "extremely powerful," a single vendor for voice, text and decisions, and enterprise controls such as data residency and Private Link that smaller providers may not offer.

Limits and open questions

The 10x speed figure is OpenAI's comparison with its own Responses API, not with rival decision models, and OpenAI has not published calibration results showing how closely gpt-6-luna's probabilities match real outcomes. For a product whose output is a probability, that is the number buyers most need.

There is still only one model, so developers cannot trade accuracy for cost within the API. Files, audio and item references are not accepted as input, and decisions that depend on an earlier answer require separate requests. The guide's price applies to /v1/decisions only; other gpt-6-luna calls follow normal pricing.

Prompt caching changes the economics only when the shared prefix is long and stable. Teams already using 12-hour prompt caching in the Responses API should check whether the same discounts apply in their Decisions workloads before moving traffic.

What to watch next

Watch for a dedicated, specifically trained decision model from OpenAI, which Handa suggested is possible, for published calibration or accuracy data, and for whether OpenAI climbs the independent decision-model leaderboards. Also watch for latency closer to the 60 milliseconds Handa mentioned, and for image outputs, which he said customers have requested.