OpenAI has opened up its Decisions API model to the general public for the first time, stepping into a fast-growing market for rapid, low-cost AI inference that recently propelled TypeSafe - a startup led by a former researcher from a leading AI lab - to a valuation of $7.6 billion.

TypeSafe's flagship Jev model effectively carved out a new category of deep learning systems that function as highly capable classifiers. Given a defined set of options, these models can make fast decisions and attach a confidence score to each outcome. Software developers quickly recognized that these so-called "decision models" or "System One models" could substitute for large language models in many workflows, delivering results significantly faster and at a fraction of the cost.

Nikunj Handa, who leads product for OpenAI's APIs, described how his team reacted upon seeing what Jev could do, noting that colleagues immediately recognized a vast range of potential applications. Rather than constructing a purpose-built decision model from the ground up, Handa's team took a different approach: they took ChatGPT Luna and restructured it to accelerate inference speeds and reduce costs.

"The big unlock with Decisions API has been the speed, shrinking things from one to three seconds to 100, 120, 180 milliseconds. In the future, we may sort of have specific training to make our models even better in these ways. But right now, it's just plain old Luna that gets the job done." - Nikunj Handa, OpenAI

From Developer Day to General Availability

OpenAI first unveiled the Decisions API at its developer day event on September 29. The product became available to all customers this past Thursday. Alongside the public launch, OpenAI introduced voice and visual capabilities to the offering, and made targeted infrastructure investments to reduce latency for users across different regions worldwide.

Early trials with enterprise customers including Box, Datadog, Ramp, and Speak gave engineers the opportunity to swap out traditional large language models for the higher-throughput Decisions API. Andrew Hsu, the chief technology officer at Speak - a company that develops AI-powered language tutoring tools - reported that the API cut the time between in-app actions from 1.4 seconds down to just 0.13 seconds. OpenAI also applied the technology internally to sharpen the responsiveness of its Dots voice assistant.

A Contested Market With No Clear Benchmark Standard

Despite the growing excitement around decision models, there is still no universally accepted measure of what makes one model better than another. Performance tends to depend on the specific use case, and the balance between intelligence, cost, calibration, and speed varies considerably by workload. TypeSafe CEO Diogo Almeida has expressed skepticism toward standardized benchmarks, encouraging developers to test models against their own real-world tasks. That said, TypeSafe itself released its own ranking system last week alongside the launch of its latest Jev model, which now includes support for image inputs and smart-if statements.

Widely referenced public benchmarks, including the AIM System One Index and the Artificial Analysis Decision Index, currently lean in favor of D1 Ultra - a new decision model released last week by Liquid - and Base Lab's Elephant-1. The newest Jev model remains competitive, placing seventh and third respectively across those two indices. OpenAI's decision model ranks eighth and fourteenth on the same benchmarks. Despite those standings, TypeSafe's products currently attract the highest query volume on OpenRouter, the multi-model API aggregation platform.

PacMan as a Proving Ground

Among the more unconventional ways to evaluate decision models is JevMan, a benchmark that tasks the models with playing the classic arcade game PacMan. The latest version of the Jev model currently tops that particular ranking. Felix Wunderlich, chief growth officer at Opper AI, the company behind the benchmark, explained why the game serves as a meaningful test. The continuous, high-pressure nature of PacMan - where a model must determine its next move within a few hundred milliseconds at every junction - creates conditions that closely mirror real-world decision-making under time constraints.

Where OpenAI's Model May Have an Edge

One area where OpenAI's offering could hold a meaningful advantage is multi-modal capability. Handa pointed to image classification as a particular strength, noting that Luna has demonstrated stronger calibration performance in that domain compared to competing models. He also mentioned that enterprise customers are pushing for the Decisions API to support image outputs, not just image inputs - a feature direction that OpenAI appears to be actively considering.

Looking further ahead, Handa sees continued speed improvements unlocking entirely new categories of applications. He raised the possibility of pushing response times down to around 60 milliseconds, which could open the door to frame-by-frame video analysis at the API level - a capability that would represent a significant leap in real-time AI processing.

Jevons' Paradox and the Road Ahead

The name Jev is itself a nod to a well-known concept in economics. The 19th-century economist William Stanley Jevons observed that when a resource becomes more efficient to use, consumption of that resource tends to increase rather than decrease - a principle now known as the Jevons Paradox. The same dynamic appears to be playing out in the AI inference market: as decision models become faster and cheaper, they are likely to be deployed in a far broader range of applications than previously imagined.

Handa echoed that view, suggesting that decision models are surfacing a growing collection of use cases that were always feasible in theory but remained impractical until now. As the technology matures and latency continues to fall, the range of problems that these lightweight, high-speed models can address is expected to expand considerably across industries and software products alike.