A Small Set of Choices Makes Model Decisions Nearly 10 Times Faster
OpenAI’s Decisions API is designed for applications that need a quick choice from a limited set of actions, rather than an open-ended response. In the launch demonstration, Romain Huet and Charlie Guo show how it turns text or image inputs into decisions an app can act on, such as routing sales leads or choosing a lane to steer a car. Huet says narrowing the model’s task to a few choices makes the API nearly 10 times faster.
A small answer space is the source of the speed
The Decisions API is designed for tasks where an app needs a model to choose among a limited set of actions quickly. As Romain Huet describes it, the developer supplies text or an image, defines a question and possible answers, and uses the model’s choice to determine what the app does next.
The API runs on GPT-6 Luna. Huet says narrowing the model’s task to a few choices makes it nearly 10 times faster while retaining image understanding, broad language support, and safety protections. The point is not to replace open-ended reasoning in every situation, but to make bounded decisions feel responsive enough to use inside an interaction.
? charlie-guo demonstrates the text path with a form-filling example. A pasted order paragraph includes a contact, company, email, product, quantity, delivery date, address, and a request for a blue cover. The app’s form has matching fields, which the API can populate from the unstructured message. Guo then extends the example: the same kind of input can be classified and routed, such as sending support tickets or inbound sales requests to the appropriate teams.
In the sales dashboard, messages from a ten-person design agency, a software company considering a 300-person rollout, and an individual freelancer are presented for classification. The interface organizes each lead around details such as company type, team size, buying stage, requirements, timeline, and requested next step. Huet says the demonstration is running at normal speed; the status readout shows 81 milliseconds of API processing for six decisions.
Visual decisions can trigger immediate actions
The image-input demonstration turns the same pattern into control. A driving game provides frames of a three-lane road with obstacles, and the API chooses a lane. The available choices are represented as lanes 1, 2, or 3; Huet says that when the game speed increases, the API still steers the car to the correct lane in a fraction of the time a reasoning model would take.
Huet suggests a related use for basic computer interactions: take screenshots, make a quick decision, and send an action back. He also draws a boundary around that example. For sophisticated computer- or browser-use scenarios, he says developers would still want Astra capabilities. The Decisions API is presented as a fast choice mechanism, not as a substitute for every more complex interaction.
The demos make the design constraint visible: the application supplies a question and a small action set, while the model interprets the input well enough to select among them. That can be useful when a slower, broader response is unnecessary—but the question and possible choices are part of how the application directs the decision.
Voice and robotics add decisions to the interaction
Guo pairs the API with GPT-Live-1 in an animated-character demo. GPT-Live-1 handles the spoken conversation; the Decisions API selects an expression from a set provided by the app. In the exchange, the character responds to Guo saying he has had a “crazy week,” then adjusts when he clarifies that it is crazy in a good way because the API is launching. Guo’s point is that an expression can add a small reaction alongside the assistant’s spoken words, rather than replacing the conversation.
Huet’s robotics demonstration uses a preview unit of Hugging Face’s programmable Microduck robot, named Lavender. GPT-Live-1 supplies voice intelligence, while the Decisions API uses the robot’s camera to choose where its head should turn. When told to “follow the apple,” Lavender turns toward the apple. Huet says the API analyzes camera frames every few moments and uses them to decide where the robot should look.
The demonstration then changes the instruction and the objects. Asked to follow “the fruit,” Lavender tracks the apple. Shown an apple and a game controller and asked which is more fun to play with, it turns toward the controller. These examples show the API selecting an action in response to spoken instructions and camera input; they do not establish how it would handle other ambiguous requests.
Across the examples, the API’s role stays consistent: translate text, images, or spoken interaction into a fast choice that another part of the application can act on. Huet presents that pattern as a basis for interactions ranging from routing a lead to steering a character or robot.