OpenAI Fully Opens Decisions API: Millisecond-Level Decision-Making Based on GPT-6 Luna Targets Fast, Low-Cost AI Market
OpenAI has officially opened its Decisions API to all customers. Optimized based on GPT-6 Luna, the interface reduces response times for certain tasks to 100-180 milliseconds, aiming to lower AI decision-making costs and accelerate agent application deployment.
According to a report by TechCrunch cited by Zhidongxi, OpenAI officially opened its Decisions API to all customers on October 22. The interface is optimized based on the reasoning model GPT-6 Luna and primarily targets tasks with defined answer ranges, such as content classification, request routing, and determining the next action of an AI agent. Developers can pre-set questions and optional answers, from which the model outputs probabilities, options, or scores.
Nikunj Handa, Head of Product for OpenAI's API, stated that after observing Jev, a decision model launched by AI startup TypeSafe, the team realized the significant potential application scenarios for such models. Instead of training a dedicated decision model from scratch, OpenAI adjusted the inference method of the existing base reasoning model GPT-6 Luna to significantly shorten response times and reduce costs. Handa noted that response times for some tasks have been reduced from 1-3 seconds to 100ms, 120ms, or 180ms.
In terms of functional specifications, the Decisions API supports text, image, and structured data inputs, handling up to 400 questions per request—double the limit during the previous public testing phase. OpenAI announced the interface at its developer day event on September 29, began public testing on October 6, and declared full availability on October 22.
OpenAI showcased several enterprise customer use cases, claiming they demonstrate the interface's advantages in speed and cost. Cloud monitoring platform Datadog uses the Decisions API to assess risks of commands generated by AI agents, reporting P95 latency approximately 40 times faster than available frontier models; financial management platform Ramp reported a 4x increase in processing speed and a 30% reduction in costs; language learning platform Speak reduced wait times between user operations from 1.4 seconds to 0.13 seconds. It should be noted that these performance figures are self-reported results from customers showcased by OpenAI and have not been verified by independent third-party large-scale benchmark reproductions.