OpenAI has opened its Decisions API to all customers, expanding the options for developers who need software to choose among alternatives quickly. The model is designed to handle decision tasks with faster responses and lower costs than using a large language model for every step.
A different role for AI models
Decision models are built to evaluate a set of options and return a choice with a confidence rating. Developers have found that this kind of system can take over some workflows that would otherwise use a large language model, particularly when speed and throughput matter.
TypeSafe’s Jev helped bring attention to the category, which is also called System One models. OpenAI’s product team saw possible applications across many workflows and built its own offering by reconfiguring ChatGPT Luna to reduce inference time and cost, rather than training a model specifically calibrated for decisions.
OpenAI API product lead Nikunj Handa described the change in response time as a major improvement: tasks that took one to three seconds could now take 100, 120, or 180 milliseconds. He said the company may pursue specific training to improve decision performance later, but the current version uses Luna.
Availability and early customer results
OpenAI announced the Decisions API at its developer day on September 29 and made it available to all customers on Thursday. The public version includes voice and visual modalities. OpenAI also said it directed inference resources geographically to reduce latency in regions around the world.
Early trials included Box, Datadog, Ramp, and Speak. Engineers at those companies used the API to replace some OpenAI large language model workflows and increase throughput. For Speak, an AI language tutor company, the change shortened the time between actions from 1.4 seconds to .13 seconds, according to CTO Andrew Hsu.
OpenAI has also used the product to make its Dots voice assistant more responsive. These examples illustrate why a specialized decision model could be useful: when a system needs to take repeated actions, shorter waits can make the overall experience feel more immediate.
Performance depends on the task
There is no single agreed measure of what makes a good decision model. The right balance among intelligence, cost, calibration, and speed can vary by workload, so benchmark rankings may not tell a developer which model will work best in a particular application.
TypeSafe CEO Diogo Almeida has urged users to test models in their own work, even as TypeSafe released its own rankings. Public comparisons such as the AIM System One Index and the Artificial Analysis Decision Index tend to rank D1 Ultra, released by Liquid, and Base Lab’s Elephant-1 highly. The latest Jev model placed seventh and third in those indices; OpenAI’s Decision Model placed eighth and fourteenth, respectively.
Those rankings are only one view of the market. TypeSafe’s models are currently the most widely queried on OpenRouter, while the JevMan benchmark tests models by having them play PacMan. The latest Jev model tops that ranking. The game tests how a model handles a continuing series of quick decisions at each junction.
More speed could widen the uses
OpenAI says its model may have an advantage in multimodal tasks. Handa pointed to image classification and said Luna has shown better calibration. Customers have also asked for more flexibility in using images as outputs, suggesting that the API’s role may extend beyond choosing among text-based options.
Further speed improvements could open another set of applications. Handa described the possibility of reaching 60 milliseconds, which could make an API call for every frame in a video imaginable. That is a potential direction, rather than a capability the article says is available now.
As decision models become faster and cheaper, developers may find more places to use them in products and software workflows. The central question for teams is practical: whether a model’s speed, confidence, and cost fit the decisions their systems need to make.