OpenAI has officially opened access to its Decisions API model, a move that puts it squarely in competition with TypeSafe’s Jev. This launch aims to serve use cases that demand fast, affordable decision-making capabilities—tasks that frequently are more efficiently handled by decision models rather than full-blown LLMs.
TypeSafe’s Jev models are designed to act as “super classifiers.” Given a set of options, they decide quickly and provide confidence metrics—serving as low-cost, high-speed alternatives to traditional language models. Their agility has already fueled high valuations and widespread industry interest.
What OpenAI Did Differently
Instead of building a bespoke decision model from scratch, OpenAI repurposed its Luna model—originally developed for ChatGPT—and reconfigured it to deliver rapid inference and reduced costs. According to internal sources, the inference time dropped from one to three seconds down to between 100 and 180 milliseconds. The company also added voice and vision capabilities, and deployed of inference resources in more regions to cut latency globally.
Since its internal debut at OpenAI’s developer day on September 29, the Decisions API has now been made generally available. Early adopters include companies like Box, Datadog, Ramp, and Speak. Speak, which builds AI-powered tutoring systems, reported an improvement in response times from approximately 1.4 seconds to just 0.13 seconds. OpenAI’s own voice assistant, Dots, has also benefitted in responsiveness.
Benchmarking the Competition
The decision-model space is already crowded. TypeSafe recently published its latest Jev model, which adds features like image input and smart-if logic. Independent benchmark indices—such as AIM System One Index and Artificial Analysis Decision Index—rate OpenAI’s Decisions API lower: its standing in these rankings places it in eighth and fourteenth spots respectively, behind newer entrants like Liquid’s D1 Ultra and Base Lab’s Elephant-1 models.
However, on the zero-shot tasks mapped for JevMan—an environment posing fast, continuous decisions modeled after classic video games—TypeSafe’s Jev still leads. JevMan emphasizes traits like speed and calibration in rapid-fire scenarios. For now, TypeSafe remains the most queried decision-model provider on OpenRouter, a public marketplace for AI models.
OpenAI believes multi-modal capacity could tip the scale in its favor. Voice and image capabilities are already integrated in Decisions API, with future expansion into image-based outputs being a priority. Product leads are also aiming to push inference times even lower—potentially toward 60 milliseconds per decision—unlocking use cases such as frame-by-frame analysis for video.
To this end, what makes a “good” decision model remains open for debate. Success depends on application: some prioritize cost, others calibration, intelligence, or raw speed. Leaders in the space advocate benchmarking models in real-world workflows rather than solely relying on labs or rankings.
Why it matters: Decision models could redefine how many tasks are handled in AI—stretching beyond chatbot responses to areas like classification, routing, and real-time decisioning. As OpenAI throws its weight behind Luna-based decision models, expect pricing, speed, and multi-modal prowess to become central battlegrounds. Keep an eye on how quickly performance improves, whether latency can meet edge demands, and whether benchmarks translate into real-world impact.
