OpenAI officially launches Decisions API with voice and image support

OpenAI has made its Decisions API generally available, offering developers a faster and lower-cost alternative to using large language models for routine classification and choice tasks.

Tim Fernholz reports for TechCrunch that the API is OpenAI’s response to a growing market for so-called decision models, also known as System One models. These systems select from defined options and return a confidence score. They can handle tasks such as sorting requests, routing content, or identifying items in images without generating long text responses.

OpenAI based the product on ChatGPT Luna, rather than training a dedicated decision model from scratch. Nikunj Handa, OpenAI’s product lead for APIs, says the main gain is speed. According to him, the API can reduce response times from one to three seconds to roughly 100 to 180 milliseconds.

The public release adds voice and image inputs. OpenAI has also placed inference capacity closer to users in different regions to reduce delays. The company says early customers including Box, Datadog, Ramp, and language-learning company Speak have used the API to increase throughput.

Speak CTO Andrew Hsu says the company reduced the time between actions in its AI tutoring product from 1.4 seconds to 0.13 seconds. OpenAI also uses the product to improve the responsiveness of its Dots voice assistant.

Competition remains intense

OpenAI enters a field shaped by TypeSafe’s Jev models, which helped popularize the category. Other contenders include Liquid’s D1 Ultra and Base Lab’s Elephant-1. Public rankings do not show a single clear leader. Fernholz notes that OpenAI’s model ranks differently across the AIM System One Index and the Artificial Analysis Decision Index.

For content and product teams, the key distinction is practical: decision models are designed for rapid, repeatable calls, while general-purpose language models remain better suited to drafting, analysis, and open-ended conversations. OpenAI says further latency reductions could eventually support decisions on every frame of a video.