Tech • AI • Robotics • Game
Introducing the Decisions API
9.2/10TL;DR
OpenAI introduced the Decisions API, a low-latency interface on GPT-6 Luna for routing text or image inputs into predefined actions in under a tenth of a second.
KEY POINTS
Fast classification for app decisions
The Decisions API is designed for cases where software must choose an action almost instantly, such as routing a request, classifying an image, or selecting an agent’s next move. Developers provide text or image input, define a question and a fixed set of possible answers, and use the returned choice directly in an application.
Built on GPT-6 Luna
The system runs on GPT-6 Luna and focuses the model on a small set of options rather than open-ended generation. That narrower task design makes it nearly 10 times faster, while retaining capabilities such as image understanding, broad language support, and built-in safety protections.
Sub-100 millisecond responses
In demonstrations involving unstructured text, the API was used to extract structured information and classify incoming messages into categories such as support or sales. In one example involving sales-lead classification, server-side responses arrived in less than 100 milliseconds, making the interaction appear effectively instantaneous.
Useful for ticketing and lead routing
A central use case is operational triage. Long, loosely written user submissions can be converted into form-like outputs or routed to the correct internal team, allowing companies to automate support queues and inbound sales sorting without waiting for a full reasoning pass.
Real-time image decisions
The API also supports visual input for rapid action selection. In a driving-style game demonstration, it analyzed road frames and obstacles, then chose whether a car should stay in lane, move left, or move right. The model continued to steer accurately even as the game speed increased, while operating at a fraction of the latency of a reasoning model.
Potential for computer control tasks
The visual decision workflow points to lightweight computer-use applications, where screenshots can be interpreted and converted into immediate actions. For more advanced browser or computer-control scenarios, richer systems may still be needed, but the new API is positioned as a fast layer for simple, repeated decisions.
Voice and animation pairing
Combined with GPT-Live-1, the Decisions API was used to control an animated character’s facial expressions during a live voice conversation. While the voice model handled spoken dialogue, the Decisions API selected from a predefined set of expressions in real time, adding visual reactions that tracked the emotional tone of the exchange.
Robotics and camera-guided actions
In a robotics demonstration, a programmable Microduck robot named Lavender used GPT-Live for voice and the Decisions API for camera-based head movement. The robot analyzed frames from its camera every few moments and chose where to look based on commands such as following an apple.
Handling abstract and ambiguous prompts
The robot was also shown responding to broader concepts rather than a single named object. When asked to follow the fruit, it tracked the apple. When asked what was “more fun to play with” between an apple and a gamepad, it selected the gamepad and continued tracking it as positions changed, illustrating concept-level classification rather than simple object matching.
CONCLUSION
The Decisions API targets a growing class of AI applications where speed matters more than long-form reasoning. By constraining outputs to defined choices, OpenAI is positioning GPT-6 Luna as a practical tool for real-time routing, visual control, and responsive interactive systems.
Ask a question