Quick Overview
This product demonstration video features Romain Huet and Charlie Guo from OpenAI introducing the Decisions API. The presentation follows recent developments in multimodal intelligence and illustrates practical use cases across structured text parsing, vision-guided navigation, and robotics.
Key Points
- 1.OpenAI introduced the Decisions API, an interface powered by GPT-6 Luna designed for rapid decision-making across text and vision inputs.
- 2.By constraining the model to a defined question and a discrete set of possible choices, the Decisions API operates nearly ten times faster than standard models.
- 3.The API processed text classification tasks, such as structuring sales leads from raw messages, in 81 milliseconds.
- 4.Vision demonstrations showed real-time lane navigation in a driving game and real-time object tracking on a programmable robot named Lavender.
- 5.When paired with GPT Live 1, the Decisions API can dynamically select expressions for animated avatars or drive physical robotic head movements alongside live voice conversations.
Summary
Romain Huet and Charlie Guo from OpenAI introduce the Decisions API, a tool engineered to deliver low-latency classification for applications requiring rapid choices. The API operates on GPT-6 Luna and functions by taking text or image inputs, a specific question, and a predefined list of potential choices. By limiting the response scope to these predefined answers, the model executes nearly ten times faster than traditional configurations while retaining standard multimodal safety, multilingual support, and image comprehension capabilities.
Charlie Guo demonstrates the API on unstructured text by automatically parsing raw messages into structured order forms and routing inbound sales inquiries. In a live interface demonstration, the Decisions API analyzes inbound customer messages to extract company size, buying stage, requirements, timeline, and next steps. The server-side processing completes six decisions in 81 milliseconds, responding well under the 100-millisecond mark.
Romain Huet then shows the API applied to visual input via a lane-navigation driving game. The Decisions API receives camera frames of the road, identifies oncoming barriers, and selects the correct lane from three numbered choices. As the simulation speed increases, the model maintains vehicle steering at speeds significantly faster than those achieved by full reasoning models. Huet notes that this low-latency vision classification can be paired with computer-use workflows alongside more complex tools like Astra.
The demonstration expands to multimodal integrations combining the Decisions API with GPT Live 1. In one example, GPT Live 1 handles spoken conversation while the Decisions API analyzes user dialogue to update the facial expressions of an animated frog avatar in real time. In a robotics demonstration using a Hugging Face Microdock prototype unit named Lavender, the Decisions API analyzes camera feeds to guide the robot's physical head movements, successfully tracking an apple and later prioritizing a game controller based on natural language queries.
Overview and Architecture of the Decisions API
Romain Huet introduces the Decisions API, powered by GPT-6 Luna, which allows applications to obtain AI classifications in a fraction of a second. By narrowing the model's focus to a specific question and a set of predefined choices, the system achieves execution speeds nearly ten times faster while maintaining image understanding, broad language support, and safety guardrails.
Text Processing and Sales Lead Routing
Charlie Guo demonstrates how the Decisions API extracts structured form fields and classifies unstructured user messages into sales lead categories. The live interface processes multiple classifications simultaneously, completing six distinct decisions and readying draft details in 81 milliseconds.
Real-Time Vision and Driving Game
The hosts demonstrate vision capabilities using a simulated driving game where the model analyzes road frames to select open lanes and avoid obstacles. The Decisions API continuously steers the vehicle at high speeds in a fraction of the time required by full reasoning models, illustrating potential for fast computer-use tasks alongside broader frameworks like Astra.
Voice Modalities and Robotics Integration
The API is demonstrated alongside GPT Live 1 to coordinate live multimodal interactions. Examples include selecting real-time emotional expressions for an animated frog character named Pond during conversation, and directing the physical head orientation of Lavender, a Hugging Face Microdock robot tracking visual objects like apples and game controllers.
The Bottom Line
The presentation establishes the Decisions API as a high-speed inference solution for constrained classification across text, vision, voice, and physical robotics. It demonstrates that limiting model outputs to discrete decision sets enables sub-100-millisecond responses without sacrificing core multimodal capabilities. The video focuses primarily on introductory demonstrations, leaving detailed documentation, pricing tiers, and public deployment timelines for subsequent developer announcements.
FAQ
What is the Decisions API and how does it function?
The Decisions API is an OpenAI tool powered by GPT-6 Luna that takes text or image inputs alongside a specified question and a set of predefined choices to produce low-latency classification outputs in a fraction of a second.
Which underlying model powers the OpenAI Decisions API demonstrations in this video?
The Decisions API runs on the GPT-6 Luna model.
How fast does the Decisions API process decisions during the inbound sales lead demo?
The Decisions API processed six distinct lead-classification decisions in 81 milliseconds on the server side.
How does the Decisions API integrate with GPT Live in interactive applications?
GPT Live handles real-time voice conversations while the Decisions API evaluates context to drive auxiliary actions, such as changing an avatar's facial expressions or steering a robot's physical gaze.
Which robotic hardware was used to demonstrate visual tracking with the Decisions API?
The demonstration used a preview unit of the programmable Microdock robot from Hugging Face, nicknamed Lavender.
Worth watching for
Software engineers, AI developers, and robotics researchers seeking low-latency classification, fast vision processing, and responsive multimodal control.
- openai
- decisions-api
- gpt-6-luna
- robotics
- gpt-live
- computer-vision