OpenAI has made its Decisions API available to all customers, expanding its push into low-latency AI models designed to make narrowly defined choices rather than generate lengthy responses. The public release adds voice and visual capabilities to a product competing with TypeSafe’s Jev and a growing field of specialized decision models.
The API, announced at OpenAI’s developer day on September 29, became broadly available on October 22. OpenAI has also distributed inference resources geographically to reduce response times for customers in different regions.
Decision models, sometimes called System One models, typically select from a supplied set of options and provide a confidence estimate. Developers can use them in workflows that otherwise require a general-purpose large language model, with speed and cost often taking priority over the ability to produce an elaborate answer. TypeSafe’s Jev helped establish demand for this approach.
OpenAI took a different technical route from building a separately calibrated decision model. Nikunj Handa, the company’s API product lead, told TechCrunch that the team adapted ChatGPT Luna to reduce inference time and expense. He said more specialized training could follow, but the current service uses the existing model.
Handa described response times falling from between one and three seconds to roughly 100 to 180 milliseconds. Those figures illustrate the product’s intended operating range, rather than a guarantee for every workload: actual application latency also depends on factors such as network travel time, input size and the surrounding software.
Early customers included Box, Datadog, Ramp and Speak, whose engineers used the API to replace existing OpenAI model calls and increase throughput. Speak, which develops AI language tutors, reported reducing the interval between actions from 1.4 seconds to 0.13 seconds. OpenAI also used the API to improve the responsiveness of its Dots voice assistant.
The launch comes as competing suppliers expand their offerings. TypeSafe’s latest Jev release added image support and smart-if statements, while Liquid introduced its D1 Ultra decision model last week. Base Lab’s Elephant-1 is another contender in public evaluations.
Benchmark results do not establish a single leader across all uses. The AIM System One Index and Artificial Analysis Decision Index tend to favor D1 Ultra and Elephant-1. OpenAI’s model placed eighth and fourteenth on those respective indices, while the latest Jev placed seventh and third. TypeSafe’s offerings nevertheless remained the most frequently queried among these competitors on OpenRouter.
For developers, the evaluation problem extends beyond choosing the fastest model. Accuracy, price and calibration—the degree to which a model’s confidence matches its actual reliability—can matter differently across applications. A system that confidently makes the wrong choice may be less useful than one that reliably signals uncertainty, particularly when software uses confidence scores to decide whether to seek human review.
OpenAI is also positioning multimodal capabilities as a point of differentiation. Handa told TechCrunch that Luna performs strongly on image classification and calibration, and that customers have asked for more flexibility around image outputs. He identified further latency reductions as a development goal, raising the possibility of decisions fast enough to process individual video frames. That remains a prospective use case rather than an announced capability.
Sources: TechCrunch AI, OpenAI