OpenAI Opens Decisions API to All Customers as Decision-Model Race With TypeSafe’s Jev Heats Up

What a decision model actually does

Decision models, sometimes called System One models after the fast, intuitive mode of human thinking, behave like super classifiers. Handed a set of options, they pick one and return a confidence rating for the choice. Developers discovered they could swap out large language models in many workflows and get answers far faster at dramatically lower cost, which created a new market segment almost overnight.

OpenAI's route into that segment was pragmatic rather than bespoke. Nikunj Handa, the product lead for OpenAI's APIs, said colleagues looked at Jev and saw a million use cases. Rather than train a purpose-built, calibrated decision model, his team reconfigured ChatGPT Luna to cut inference time and cost. "The big unlock with Decisions API has been the speed, shrinking things from one to three seconds to 100, 120, 180 milliseconds," Handa said, adding that the current version is plain Luna doing the job, with dedicated training a possible future step.

Speed as the selling point

The public release adds voice and visual modalities and targets lower latency worldwide by placing inference resources geographically. Early trials with Box, Datadog, Ramp and Speak showed engineers replacing OpenAI's LLMs to gain throughput. Andrew Hsu, CTO of Speak, which builds AI language tutors, said the API cut the gap between actions from 1.4 seconds to 0.13 seconds. OpenAI also used the product to sharpen the responsiveness of its Dots voice assistant.

Handa sees further headroom. Pushing latency toward 60 milliseconds would make a call on every frame of video conceivable, opening use cases that were previously impractical. That trajectory echoes the economist after whom Jev is named, whose work described how using a resource more efficiently makes it more ubiquitous.

A crowded and unsettled benchmark landscape

What makes a good decision model is still unsettled beyond the practical test of whether it works for a given workload. Optimization across intelligence, cost, calibration and speed appears to vary by task. TypeSafe CEO Diogo Almeida distrusts benchmarks and urges teams to trial models on their own work, though his company published its own rankings last week alongside its latest Jev model, which adds image support and smart-if statements.

Public indices such as the AIM System One Index and the Artificial Analysis Decision Index currently favor D1 Ultra, a new decision model Liquid released last week, and Base Lab's Elephant-1. The latest Jev sits seventh and third on those two indices, while OpenAI's decision model lands eighth and fourteenth. TypeSafe's offerings remain the most widely queried on OpenRouter. There is also JevMan, a benchmark in which decision models play Pac-Man; the latest Jev tops it. Felix Wunderlich, chief growth officer at Opper AI, which built the benchmark, called the game an interesting test because of its nonstop stream of quick, high-stakes decisions, with the model choosing a direction at every junction in a few hundred milliseconds.

Where OpenAI thinks it can win

OpenAI's advantage may lie in multimodality. For image classification, Handa said Luna is extremely powerful and has shown better calibration, and customers want the Decisions API to become more flexible about using images as outputs as well as inputs. That matters as decision models move beyond text routing and into workflows where visual context is part of the choice.

The broader contest now pits a startup that defined the category against a platform vendor that can repurpose an existing model and distribute it through an established API business. As latency falls and prices follow, the expectation is that decision models will surface use cases that were hiding in plain sight, expanding the market rather than simply dividing it.