OpenAI Launches Decisions API to Challenge Specialized AI Models

The shift toward decision models marks a departure from traditional large language models, favoring speed over verbose generation. For software developers, these tools act as super classifiers that process options and provide confidence ratings in milliseconds rather than seconds. OpenAI product lead Nikunj Handa noted that the team adapted their existing Luna model to achieve latency between 100 and 180 milliseconds, a significant reduction from the standard one-to-three-second window required by earlier iterations.

Early adopters, including Box, Datadog, Ramp, and language tutor platform Speak, have already integrated the API to streamline operations. Speak reported cutting action times from 1.4 seconds to just 0.13 seconds. While competitors like TypeSafe and Liquid currently lead on various industry benchmarks—such as the AIM System One Index and the quirky PacMan-based JevMan test—OpenAI is banking on its multimodal capabilities. Handa pointed to Luna’s superior image classification as a primary differentiator, suggesting that further latency optimization could soon allow for decision-making on a per-frame basis in video processing.